Freezing of Gait in Parkinson's Disease: A Scoping Review on the Path Towards Real-Time Therapies.

Freezing of Gait in Parkinson’s Disease: A Scoping Review on the Path Towards Real-Time Therapies - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sensors (Basel) . 2026 Mar 25;26(7):2042. doi: 10.3390/s26072042 Search in PMC Search in PubMed View in NLM Catalog Add to search Freezing of Gait in Parkinson’s Disease: A Scoping Review on the Path Towards Real-Time Therapies Meenakshi Singhal Meenakshi Singhal 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) Find articles by Meenakshi Singhal 1 , Christina Grannie Christina Grannie 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) Find articles by Christina Grannie 1 , Margaret Burnette Margaret Burnette 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) Find articles by Margaret Burnette 1 , Manuel E Hernandez Manuel E Hernandez 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) 2 Department of Bioengineering, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA 3 Beckman Institute for Science and Technology, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA Find articles by Manuel E Hernandez 1, 2, 3, * , Samar A Hegazy Samar A Hegazy 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) Find articles by Samar A Hegazy 1, * Editor: Christian Peham Author information Article notes Copyright and License information 1 Carle Illinois College of Medicine, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA; ms189@illinois.edu (M.S.); grannie2@illinois.edu (C.G.); phburn@illinois.edu (M.B.) 2 Department of Bioengineering, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA 3 Beckman Institute for Science and Technology, University of Illinois at Urbana-Champaign, 506 South Mathews Avenue, Urbana, IL 61801, USA * Correspondence: mhernand@illinois.edu (M.E.H.); hegazy@illinois.edu (S.A.H.); Tel.: +1-(217)-224-8971 (M.E.H.); +1-(217)-333-0979 (S.A.H.) Roles Christian Peham : Academic Editor Received 2025 Dec 31; Revised 2026 Mar 6; Accepted 2026 Mar 20; Collection date 2026 Apr. © 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license . PMC Copyright notice PMCID: PMC13074614  PMID: 41977827 Abstract Background: Freezing of gait (FoG) is a common symptom of Parkinson’s disease, especially in its later stages of progression. Characterized by involuntary stopping during normal gait patterns, FoG greatly increases fall risk, reducing quality of life. Given the complex presentation and etiology of FoG, current treatments have proven ineffective in managing episodes. In recent years, machine learning algorithms have been leveraged to derive actionable clinical insights from biomedical datasets. As a manifestation of neuromechanical dysfunction, impending FoG episodes may be characterized through data collected by wearable devices and sensors. Objective: This scoping review evaluates the current landscape of machine and deep learning-derived biomarkers to enhance the personalized management of FoG. Methods: This scoping review was conducted using established methodological frameworks for scoping reviews and is reported in accordance using the PRISMA-ScR checklist. Three databases were queried, with screening yielding 60 studies. Results: Thirty-nine papers reported on deep learning techniques, with the most common architectures being convolutional neural networks and long short-term memory models. Conclusions: Inertial measurement units, which can be worn on various locations, may be a promising modality for practical implementation. To generate closed-loop FoG therapies, algorithms can be integrated into real-time systems like robotic exoskeletons or adaptive deep brain stimulation. Future work in generating datasets from ambulatory devices, as well as distributed computing strategies, may lead to real-time FoG management. Keywords: closed loop, machine learning, Parkinson’s disease, freezing of gait 1. Introduction Parkinson’s disease (PD) is a neurodegenerative movement disorder affecting over 1 million people in the U.S. and 10 million worldwide [ 1 ]. Clinical features include both motor and non-motor symptoms, such as sleep disturbances, fatigue, depression, anxiety, and apathy. Motor symptoms include tremors, shuffling, bradykinesia, postural instability, and rigidity, which can all drastically reduce quality of life when inadequately managed. In particular, freezing of gait (FoG) is a phenomenon where patients temporarily lose the ability to move their legs forward while walking. Although FoG is most associated with advanced disease progression, it has been reported in approximately 40% of the overall Parkinson’s disease population and in more than 70% of patients with a disease duration of 10 years or longer [ 2 ]. Often triggered by sudden movement changes like turning, FoG episodes significantly increase the risk of falls, making them a critical area within PD. A major clinical challenge in FoG management is its variable presentation across patients. While some may experience gait freezing when they first initiate movement from rest, others may find they are more susceptible to episodes while dual-tasking, such as both speaking and walking. Moreover, differences in the nature of festination, i.e., foot shuffling, that patients express prior to freezing create another parameter of consideration [ 3 ]. Standard therapies, including levodopa regimens [ 3 ] and continuous deep brain stimulation (DBS) [ 4 ], have demonstrated limited efficacy in mitigating symptoms. Certain patients may be more susceptible to FoG during medication ON state, OFF state, or transition phases between medication states; meanwhile, there may be a trade-off between gait-optimized DBS frequency and appendicular side effects [ 5 ]. Common triggers include turning, doorway/narrow-space negotiation, and dual-tasking, contributing to between-patient variability in freezing phenotypes. While the underpinnings of FoG remain an area of active study, the neuromechanical nature of its presentation lends itself to a variety of analytical techniques. The rise in wearable sensor technologies within healthcare has facilitated a revolution of continuous monitoring, rehabilitation, and other applications that leverage data-driven approaches. In particular, the development of digital biomarkers derived from inertial measurement units (IMUs) has enabled improved management of neuromuscular conditions, including stroke [ 6 , 7 ] and muscular dystrophy assessment [ 8 ]. Machine learning (ML) is a subset within the larger field of artificial intelligence (AI), with algorithms allowing for prediction and decision-making by learning trends from given data in a supervised manner. Traditional ML algorithms include linear regression, support vector machines (SVMs, Figure 1 A), k-nearest neighbors (kNN), tree-based methods, and the naive Bayes classifier. While each model processes information in a different manner, they all train by making decisions on pre-labeled data. Deep learning (DL) is a more specialized subset of ML in which neural network models can infer underlying trends from unlabeled data, optimizing their own actions through an unsupervised learning process. While ML algorithm performance may plateau, DL performance tends to increase alongside data input. By combining multiple sets of hidden layers and weight functions, convolutional neural networks (CNNs, Figure 1 B) and other DL architectures offer extensive flexibility for model optimization. Although DL algorithms are defined by their ability to provide pattern recognition insights in large-scale, complex datasets, there are tradeoffs in terms of training time and feature interpretability. Figure 1. Open in a new tab Representation of ( A ) support vector machine, which optimizes the position of a decision boundary called the hyperplane to perform classification and regression tasks; ( B ) CNN, whose parameters (e.g., layer number and type, filtering, activation functions) can be highly customized. By training on transcriptomic, imaging, and other data sources, ML/DL has powered the discovery of biomarkers in cancer [ 9 ], neurological conditions [ 10 ], and diabetes [ 11 ]. At the time of writing, upwards of 80% of AI algorithms currently approved by the U.S. Food and Drug Administration are within radiology [ 12 ], highlighting the ease of data acquisition in the field. For freezing of gait, recent data science contests suggest a community-wide push towards establishing clinically relevant AI models [ 13 ]. Identifying objective markers for FoG is the first step towards developing closed-loop, real-time therapies. To this end, we conducted a scoping review to map the current progress and challenges in applying ML techniques to datasets, towards enabling data-driven methods to improve the management of FoG in patients with Parkinson’s disease. 2. Materials and Methods This scoping review was conducted using established methodological frameworks for scoping reviews and is reported in accordance with the PRISMA-ScR guidelines ( Supplementary Materials ) [ 14 ]. The PubMed, Web of Science, and SCOPUS databases were queried on 4 February 2025, with a curated search strategy, with the PubMed version provided: ((artificial intelligence[majr:noexp] OR machine learning[majr:noexp] OR deep learning[majr:noexp] OR neural networks, computer[majr:noexp] OR (artificial intelligence[tiab] OR machine learning[tiab] OR deep learning[tiab] OR neural networks[tiab])) AND (Parkinson disease[MeSH] OR “Parkinson’s disease”[tiab] OR “Parkinson disease”[tiab])) AND ((“freezing of gait”[tiab])) This search strategy yielded 122 unique results ( Figure 2 ). Peer-reviewed papers in the English language were included if they: (1) were published between 2020 and 2025; (2) address patients diagnosed with PD; (3) focus on predicting, detecting, or treating FoG episodes; (4) involve the application of machine learning algorithms to process data; and (5) include analyses of datasets based on electrophysiological and/or biomechanical data. Given the rapid advancements in machine learning within healthcare and wearable devices, only papers from the given time range were included. While literature may have explored FoG within a broader analysis of mobility and Parkinson’s disease, those papers may not have been represented within the results, as the focus of the search query was on FoG. Exclusion criteria included editorials, review papers, book chapters, animal model studies, data science competitions or other formats not exploring the work generated by a single team in depth, or works not focused on the management of FoG. The full dataset independently underwent title and abstract screening by two reviewers (M.S. and C.G.), and discrepancies were resolved through discussion. Full-text review on the resulting 64 studies was split between both reviewers, who held joint discussion on decisions for all papers. Following full-text screening, 60 papers underwent data extraction. Figure 2. Open in a new tab PRISMA diagram depicting study selection process. 3. Results In general, FoG management was explored as either episode prediction, detection, or therapeutic intervention within the 60 included studies ( Table 1 ). In addition to performance measures like sensitivity, specificity, and AUC, domain-specific metrics such as latency time and computational complexity were frequently assessed to evaluate the translational potential of a given algorithm. Table 1. Characterization of n = 60 included studies; AUC = area under the curve, ECoG = electrocorticography, CNN = convolutional neural network, GBM = gradient boosting machine, IMU = inertial measurement unit, kNN = k nearest neighbor, PD = Parkinson’s disease, RNN = recurrent neural network, TCN = temporal convolutional network. Author Patient Population Machine Learning Data Modalities Model Type and Features Performance Metrics Slemenšek et al. [ 15 ] 6 PD patients deep learning accelerometer, gyroscope, and muscle activity sensors CNN + RNN 95.1% accuracy, 98.8% specificity, 2.3% precision, 2.4% sensitivity Koltermann et al. [ 16 ] 26 PD patients deep learning 2 IMUs (L/R ankles) multivariate time-series transformer encoder 96.5% accuracy Chan et al. [ 17 ] 107 PD patients deep learning 1 IMU (waist) CNN 83% accuracy, 89% sensitivity, 81% specificity Al-Nefaie et al. [ 18 ] 65 PD patients machine learning 1 IMU (lower back) decision tree 91% accuracy, 97% precision, 99% sensitivity, 0.94 F1-score Dvorani et al. [ 19 ] 16 PD patients machine learning 1 IMU (foot) AdaBoost 0.875 ± 0.082 AUC, 84.0% ± 7.0% accuracy, 85.5% ± 8.6% sensitivity, 78.8% ± 11.9% specificity Li et al. [ 20 ] 50 PD patients machine learning Mobile phone video recordings (OpenPose annotations) XGBoost 81.56% accuracy, 87.5% sensitivity, 79.82% specificity Mesin et al. [ 21 ] 12 PD patients machine learning 2 IMUs (L tibia and wrist), and EEG Support vector machine 88% accuracy, 85.14% sensitivity, 88.38% specificity, 87.71% precision, 0.8673 F1-score Liu et al. [ 22 ] 8 PD patients machine learning ECoG and local field potentials mixed-effects random forest 77% accuracy, AUC: 0.75 Borzi et al. [ 23 ] 11 PD patients machine learning 2 IMUs (L/R shins) Support vector machine 87.4% accuracy, 84.0% sensitivity, 88.3% specificity Pardoel et al. [ 24 ] 11 PD patients machine learning 4 IMUs (L/R shanks, thighs) and plantar pressure sensors Random Undersampling-boosted decision-tree ensemble model 76.4% sensitivity, 86.2% specificity Zhang et al. [ 25 ] 12 PD patients machine learning 1 IMU (lower back) AdaBoost 82.7% accuracy Demrozi et al. [ 26 ] 10 PD patients machine learning Daphnet dataset (3 IMUs: ankle, hip, knee) kNN 94.1% sensitivity, 97.1% specificity Abbasi et al. [ 27 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) Bidirectional LSTM (with ensemble learning and bottleneck attention module) 99.88% accuracy, 99.99% sensitivity, 98.89% specificity Wang et al. [ 28 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) multi-channel time-series neural network; intra/inter-channel transformer; attention mechanism 96.21% accuracy, 78.64% precision, 99.13% specificity Sun et al. [ 29 ] 45 PD patients deep learning Video recordings (OpenPose annotations) higher order (3rd-order) polynomial transformer (with self-attention mechanism) 86.1% accuracy, sensitivity: 86.6%, specificity: 84.7%, AUC: 0.92 Yang et al. [ 30 ] 12 PD patients deep learning 5 IMUs (pelvis, L/R tibias, tali) temporal CNN 0.63 F1-score Park et al. [ 31 ] 12 PD patients deep learning Pedar shoe insole system (big toe, forefoot, heel) temporal CNN 99% accuracy, 68% precision, 88% sensitivity, 90% specificity, 0.76 F1-score Klaver et al. [ 32 ] 70 PD patients deep learning combined 4 datasets (7 IMUs: lower back, L/R upper legs, lower legs, feet) CNN 85% sensitivity, 68% specificity, 0.86 AUC Borzi et al. [ 33 ] 59 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee), ADL dataset (1 IMU: lower back), REMPARK dataset (1 IMU: L waist) CNN 95.6% sensitivity, 92.9% specificity, AUC: 0.979, 0.621 F1-score Hu et al. [ 34 ] 21 PD patients deep learning Footstep pressure sensor gait mat Adversarial spatial-temporal network 75.7% accuracy, 0.847 AUC Hou et al. [ 35 ] 12 PD patients deep learning electroencephelography, 2 IMUs (L/R tibias), and EMG (L/R tibias) CNN 81% sensitivity, 88% specificity, 0.85 AUC Borzi et al. [ 36 ] 118 PD patients and 21 healthy elderly subjects deep learning REMPARK dataset (1 IMU: L waist), 6 MWT (1 IMU: lower back), ADL dataset (1 IMU: lower back), CNN 0.877 sensitivity, 0.883 specificity, 0.830 F-1 score, 0.946 AUC Kwon et al. [ 37 ] 57 PD patients deep learning 3D motion capture attention-based adaptive graphical convolutional network 0.962 F1-score Shi et al. [ 38 ] 63 PD patients deep learning 2 IMUs (L/R ankles) CNN 0.915 F1-score, 90.7% Geometric mean score O’Day et al. [ 39 ] 7 PD patients deep learning 3 IMUs (lower back, L/R ankles) CNN 0.83 AUC Filtjens et al. [ 40 ] 14 PD patients deep learning 3D motion capture CNN (with layer-wise relevance propagation) 86.8% accuracy, 82.1% sensitivity, 88.9% specificity Shi et al. [ 41 ] 67 PD patients deep learning 3 IMUs (lower back, L/R ankles) 2DCNN 89.2% accuracy, 82.1% sensitivity, 96% specificity, 88.8% Geometric mean score Pardoel et al. [ 42 ] 11 PD patients machine learning plantar pressure data feature selection (Relief-F) followed by ensemble decision tree classifiers 77.3% sensitivity, 82.9% specificity, 94.2% FOG episodes identified, 0.8 s average prediction time before FOG onset Bikias et al. [ 43 ] 11 PD patients deep learning CuPiD dataset (1 IMU: wrist) CNN 86% sensitivity, 90% specificity Mostafa et al. [ 44 ] 24 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee), A.T. Still dataset (6 IMUs: sternum, lower back, L/R wrists, ankles) LSTM + CNN Ankle data: 98.5% ± 0.3% accuracy, 98.5% ± 0.3% precision, 98.5% ± 0.3% sensitivity, 97.9% ± 0.6% accuracy Shalin et al. [ 45 ] 11 PD patients deep learning plantar pressure data LSTM FOG detection: 82.1% sensitivity, 89.5% specificity; FOG prediction: 72.5% sensitivity, 81.2% specificity Dvorani et al. [ 46 ] 16 PD patients machine learning 1 IMU (foot) Support vector machine 81.6 ± 7.7% sensitivity, 70.3 ± 18.4% specificity, 0.828 ± 0.071 AUC Miao et al. [ 47 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) deep residual network (ResNet) 85.7% sensitivity, 94.0% specificity, 89.7% Geometric mean score Pardoel et al. [ 48 ] 11 PD patients machine learning 4 IMUs (L/R lower thighs, ankles) and plantar pressure feature selection (Relief-F) followed by decision tree ensemble FOG detection: 80.2% sensitivity, 86.3% specificity, 72.6% precision; FOG prediction: 72.2% sensitivity, 83.6% specificity, 60.2% precision Sigcha et al. [ 49 ] 21 PD patients deep learning 1 IMU (waist) CNN + LSTM 87.1% sensitivity, 87.1% specificity, 87.1% Geometric mean score, 0.939 AUC Li et al. [ 50 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) LSTM (with squeeze and excite block and self-attention mechanism) 0.927 sensitivity, 0.952 specificity, 0.976 AUC Liu et al. [ 51 ] 24 PD patients machine learning 7 IMUs (waist, L/R calves, thighs, feet) Light GBM 88.62% accuracy, 87.69% sensitivity, 89.29% specificity, 0.8849 AUC Hu et al. [ 52 ] 45 PD patients deep learning video recordings Graph Sequence RNN 82.5% accuracy, 83.8% sensitivity, 82.3% specificity, 0.90 AUC Borzi et al. [ 53 ] 38 PD patients and 21 elderly subjects machine learning 1 IMU (waist) Support vector machine 98.3% accuracy, 95.4% sensitivity, 98.8% specificity, 92.8% precision Reches et al. [ 54 ] 57 PD patients machine learning 3 IMUs (lower back, L/R ankles) random forest 85.0% ± 10.0% accuracy, 84.1% ± 22.3% sensitivity, 83.4% ± 12.2% specificity, 0.93 AUC Shalin et al. [ 55 ] 11 PD patients deep learning plantar pressure data collected from pressure sensor arrays CNN 94.3% sensitivity, 95.1% specificity Filtjens et al. [ 56 ] 15 PD patients deep learning 3D motion capture TCN Initial contact model: 0.997 F1-score End contact model: 0.999 F1-score Martínez-Villaseñor et al. [ 57 ] 10 PD patients machine learning Daphnet dataset (3 IMUs: ankle, hip, knee) Artificial hydrocarbon network 88% accuracy, 88% sensitivity, 88% specificity, 0.88 F1-score Unni et al. [ 58 ] 9 PD patients machine learning stepping from force plates Random Forest Values not directly reported Ezhilarasi et al. [ 59 ] 12 PD patients deep learning 2 IMUs (L/R shins) Faster Mask Region-Based CNN 92% sensitivity, 91% specificity, 0.92 AUC, 95.3% Geometric mean score Kumar et al. [ 60 ] 122 PD patients deep learning TDCSFOG dataset (1 IMU: lower back), 1D-CNN with attention mechanism 0.8857 AUC DeFOG dataset (1 IMU: lower back) Kondo et al. [ 61 ] 16 PD patients deep learning object tracking and 3D pose estimation 1D CNN + LSTM 93.2% accuracy, 88.8% sensitivity, 97.9% specificity, 97.9% precision, 0.931 F1-score Pardoel et al. [ 62 ] 17 PD patients machine learning plantar pressure data decision tree ensemble with random undersampling boosting (RUSBoosting) 77.42% sensitivity, 80.39% specificity Yang et al. [ 63 ] 18 PD patients deep learning 5 IMUs (pelvis, L/R tali, tibias) 1D TCN 0.78 F1-score Banu et al. [ 64 ] 10 PD patients machine learning Daphnet dataset (3 IMUs: ankle, hip, knee) CNN 100% accuracy, 100% sensitivity, 1.00 F1-score Bacanin et al. [ 65 ] 93 PD patients deep learning plantar pressure data gated recurrent unit (RNN) with crayfish optimization algorithm 87.0% accuracy, 90.3% precision, 82.3% sensitivity, 86.1% F1 score Peppes et al. [ 66 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) Generative Adversarial Network 90.66% accuracy Dimoudis et al. [ 67 ] 45 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee), IMU dataset (1 IMU: leg) inception CNN with squeeze and excite block 98% sensitivity, 99% specificity, 0.972 F1-score, AUC: 0.986 Li et al. [ 68 ] 32 PD patients deep learning 2 IMU accelerometers (L/R ankles) and force-sensitive insoles LSTM + CNN with squeeze and excite block and attention mechanism feature-fusion-weighted model: 96.3% accuracy, 92.4% sensitivity, 98.3% specificity, 0.943 F1-score Önder et al. [ 69 ] 10 PD patients machine learning Daphnet dataset (3 IMUs: ankle, hip, knee) Gradient boosted decision trees 92.7% accuracy, 66.7% sensitivity, 96.6% specificity Koltermann et al. [ 70 ] 11 PD patients deep learning 2 IMU accelerometers (L/R ankles) CNN Not directly reported Sadiq et al. [ 71 ] 23 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee), PhysioNet dataset (ground reaction forces) CNN + LSTM with attention mechanism 98.74% accuracy, 99.25% precision, 99.38% sensitivity, 96.18% specificity, 0.9930 F1-score, 1.00 AUC Mekruksavanich et al. [ 72 ] 10 PD patients deep learning Daphnet dataset (3 IMUs: ankle, hip, knee) CNN with squeeze and excite mechanism 95.66% accuracy, 95.58% precision, 95.66% recall, 95.56% F1-score Ren et al. [ 73 ] 12 PD patients machine learning 7 IMU accelerometers (waist, L/R thighs, ankles, feet), plantar pressure data random forest 88.09% accuracy, 78.39% sensitivity, 91.66% specificity, 77.58% precision and 0.7798 F1-score Filtjens et al. [ 74 ] 14 PD patients deep learning 3D motion analysis multi-stage spatial-temporal graph convolutional network (MSTGCN) Not directly reported Open in a new tab 3.1. Data Modalities, Features, and Model Performances The models in the included studies were trained on several key categories of data from patients with Parkinson’s disease. Thirty-seven papers utilized inertial measurement unit (IMU) recordings containing accelerometry (and occasionally gyroscopic and/or magnetometry) data taken from various locations on the body ( Figure 3 ). Table 2 further explores the variability and trends in IMU positioning. Anywhere between one and seven sensors were placed on participating patients, with the optimal combination and number differing across studies. Several papers have demonstrated that sensor count does not correlate with higher model performance for FoG or other gait-related conditions. Rather, sensor positioning plays a larger role in effective mobility assessment and long-term monitoring [ 39 , 75 , 76 ]. From a translational perspective, the authors note that certain locations, such as the ankle, may be more feasible for incorporation within a closed-loop wearable for FoG reduction. Key considerations include noise, patient preference, and anatomical variability. Overall, the most common IMU placement was at the ankle, followed by the hips and knees. Figure 3. Open in a new tab Papers stratified by: ( A ) number of PD patients included in dataset; ( B ) type of training data, with multimodal including combinations of several forms. Table 2. Papers stratified by the location of IMUs used in the training dataset. Note: Some papers utilized multiple sensor locations. IMU Location Number of Papers References Back 10 [ 18 , 25 , 32 , 33 , 36 , 39 , 41 , 44 , 54 , 60 ] Sternum 1 [ 44 ] Waist 7 [ 17 , 33 , 36 , 49 , 51 , 53 , 73 ] Hip/Pelvis 16 [ 26 , 27 , 28 , 30 , 33 , 44 , 47 , 50 , 57 , 63 , 64 , 66 , 67 , 69 , 71 , 72 ] Thigh/Upper legs 6 [ 24 , 32 , 48 , 51 , 67 , 73 ] Knees 14 [ 26 , 27 , 28 , 33 , 44 , 47 , 50 , 57 , 64 , 66 , 67 , 69 , 71 , 72 ] Tibia/Shin/Lower legs/Shank/Calves 9 [ 21 , 23 , 24 , 30 , 32 , 35 , 51 , 59 , 63 ] Ankle/talus 25 [ 16 , 26 , 27 , 28 , 30 , 33 , 38 , 39 , 41 , 44 , 47 , 48 , 50 , 54 , 57 , 63 , 64 , 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 ] Foot 5 [ 19 , 32 , 46 , 51 , 73 ] Wrist 3 [ 21 , 43 , 44 ] Open in a new tab Some groups found that incorporating wavelet-based features was particularly beneficial for model training time and performance. With analytical methods such as the Fast Fourier transform, time-domain data can be translated into the frequency domain, revealing new temporal information. Noting the heterogeneity in FoG presentation across patients, Shi et al. [ 41 ] aimed to unite the two domains by generating time-frequency wavelet representations. The researchers converted raw time-series IMU recordings from 67 PD patients into Morlet continuous wavelet transforms, achieving an F-1 score of 91.5% using an 8-layer CNN architecture. Meanwhile, eight papers involved plantar pressure data collected from wearable insole sensors or force plates. Taking advantage of the cyclical nature of overground walking, researchers harnessed features like ground reaction forces. In general, models trained on video or 3D motion capture data ( n = 8) had lower performance across both ML and DL architectures, with accuracies in the 70–80% range. While kinematic features provide insight on the biomechanical component of FoG, its onset as dysregulation within basal ganglia circuitry allows neurophysiological information to lend another perspective. One study incorporated solely neural data in the form of electrocorticography (ECoG) recordings taken from patients with DBS implants. Liu et al. [ 22 ] found that beta (13–30 Hz) power in the subthalamic nucleus (STN) was significantly higher during FoG episodes than in normal walking, with minimal variability throughout standard walking periods. The researchers allude to modifying DBS algorithms to target FoG based on such gait-specific patterns, a concept further explored in the Discussion. Six papers employed a combination of multiple data streams, such as electroencephalography (EEG) in tandem with electromyography (EMG). Li et al. [ 68 ] proposed a form of multimodal fusion with adaptive weighting, integrating IMU and insole sensor data. The group noted that their “feature-fusion-weighted” model outperformed models trained on a single data modality. 3.2. Dataset Diversity and Augmentation To address concerns related to dataset quality, papers incorporated various data augmentation strategies. Thirty-seven of the sixty studies trained on datasets from cohorts of twenty or fewer PD patients ( Figure 3 ); 14 of these utilized the publicly available Daphnet Freezing of Gait dataset. Daphnet, curated in 2013, contains 237 FoG events recorded from ten patients diagnosed with Parkinson’s disease, with an average age of 66.5 years. A total of 8 h and 20 min of recordings are included from three triaxial accelerometry sensors (ankle, hip, and knee) worn during both assigned walking tasks and activities of daily living (ADLs) [ 77 ]. Hinting at the need for more diverse and thorough datasets in this space, one study focused exclusively on drawing from generative artificial intelligence (genAI) techniques to create a novel dataset. By training a generative adversarial network on Daphnet, Peppes et al. [ 66 ] created a synthetic FoG dataset that, when validated with a deep neural network, achieved 92.09% accuracy in FoG episode classification. Given the challenges of collecting ambulatory data, patients were generally recorded within a standardized clinic setting as opposed to an at-home environment. Thus, for the majority of studies, FoG episodes were “induced” by various exercises like the Timed Up and Go (TUG) mobility assessment or dual-tasking. In line with this trend, a considerable number of datasets were highly imbalanced, with a relative scarcity of FoG episodes relative to the total recording time. Al-Nefaie et al. [ 18 ] and Mostafa et al. [ 44 ] address this discrepancy through the “synthetic minority oversampling strategy” (SMOTE), a widely used resampling method. SMOTE functions by sampling along temporary delineations between data points in the minority class (e.g., FoG episodes) that are close in proximity within the feature space [ 78 ]. While the three teams built models with notably high performance on FoG prediction tasks ( Table 1 ), oversampling and genAI methods raise broader questions about the translation of algorithms trained on data patterns not present in the real world. 3.3. Algorithmic Design and Parameters For the 22 papers based on conventional ML algorithms, the SVM most frequently had the highest performance, followed by kNN. Though decision tree ensembles like random forest, XGBoost, and AdaBoost were the most common supervised model category, they typically presented with lower metric scores. For these supervised ML papers, ground truth was usually established by clinician-labeled FoG episodes. Thirty-nine of sixty papers applied deep learning models ( Figure 4 ) to FoG datasets; of these studies, three found long short-term memory (LSTM) models to be the most effective, and the rest reported on another form of recurrent neural network (RNN), transformer, adversarial neural network, or CNN. In general, RNNs ( Figure 5 ) function to process sequential data with built-in feedback loops, making them well-suited for time-series analysis. LSTMs build upon this framework through an added memory gating system ( Figure 5 ). Figure 4. Open in a new tab Papers stratified by: ( A ) supervised ML algorithm; ( B ) unsupervised ML algorithm. Ensemble learning models involved CNNs combined with another type. Figure 5. Open in a new tab Representation of ( A ) recurrent neural network (RNN); ( B ) long short-term memory model (LSTM). LSTMs build upon the standard RNN by incorporating an additional “forget gate”. This enables selective information retention, allowing the model to more efficiently learn long-term dependencies. 3.4. Strategies for Improved Performance Studies utilizing DL algorithms applied several strategies to improve model performance. Six papers employed ensemble learning, which combines two or more architectures to leverage their individual strengths. Each ensemble contained a CNN in conjunction with either an LSTM or RNN, allowing for multifunctional benefits. While CNNs are generally suited for spatial data, LSTMs can provide insights into time-dependent trends. Indeed, the performance of all ensemble classifiers was in the 95–99+% range, except for one that trained on video recordings. Meanwhile, teams focusing on LSTMs found that adding an attention mechanism ( Figure 6 ) into the architecture enhanced FoG episode detection. By assigning greater weights to certain features, an attention mechanism boosts a model’s ability to capture large-scale relationships and process complex time series data, as in the case of sequential recordings containing FoG episodes. The squeeze-and-excite block, one form of attention mechanism, offers versatile integration into DL architectures with minimal added computational cost. In the 60 studies, it was incorporated within both LSTMs and CNNs and could be separated into three components: (1) squeeze; (2) excite; (3) scale. During the first phase, global average pooling condenses the fully connected layer into a single embedding, improving generalizability on new data. In the latter phases, the channels are recalibrated based on the simplified embeddings [ 79 ]. CNNs were also augmented to address interpretability: Filtjens et al. [ 40 ] utilized layer-wise relevance propagation ( Figure 6 ), an Explainable Artificial Intelligence method, to explore the most salient joint trajectory features taken from 3D motion capture software. Figure 6. Open in a new tab Representation of ( A ) attention mechanism, utilized to help direct DL models towards the most relevant information; ( B ) layer-wise relevance propagation, which involves decomposing model outputs to derive the most meaningful input features. 3.5. Closed-Loop Systems Ultimately, the clinical management of FoG depends on therapeutic intervention in addition to episode detection. Six papers demonstrate “closed-loop” systems that aim to do so. In 2023, Koltermann et al. [ 70 ] published FoG-Finder, which functions with a 615 ms latency period between FoG episode prediction and cueing. FoG-Finder implements a CNN architecture trained on ankle accelerometry data to time the delivery of vibrotactile stimulation based on detected FoG episodes. Koltermann et al. [ 16 ] went on to develop Gait-Guard, which once more provides vibrotactile cueing based on ankle IMU recordings, now processed by a transformer DL model. The authors note that Gait-Guard had an improved latency period of 378.5 ms, suggesting real-world benefit to affected patients and the potential for architecture optimization to improve performance. Dvorani et al. [ 19 ] emphasize their advances in edge computing, a distributed model where data processing occurs along multiple nodes (e.g., sensors) that are closer to the user, rather than in a centralized location. This methodology enabled researchers to obtain a latency of 7.5 ms between sensor recording and electrical stimulation. Hou et al. [ 35 ] implemented multimodal prediction based on signals from a custom around-the-ear wearable EEG and flexible sensors capable of recording EMG and accelerometry signals from the calves and shins. When an impending FoG episode is detected by a pretrained CNN classifier powered by an integrated circuit, an auditory cue is delivered to the patient via a speaker. Although latency values are not reported, the researchers highlight the low power consumption of the hardware, as well as the model’s high AUC (0.85). 4. Discussion The variety of AI models that researchers tested emphasizes the clinical challenge presented by FoG. Across the included studies, DL methods demonstrated improved performance compared to traditional ML models, which may be in part due to the complex nature of overground walking. While FoG can be modeled as a binary classification problem, the episode is preceded by a series of gait events, like shuffling or trembling, that contain information about underlying disease progression, pathophysiology, and other nonlinear features that DL approaches may be more amenable towards detecting. In particular, LSTMs, which can capture trends in sequential data, could be an ideal baseline architecture to explore and augment for FoG management. The results of this review paper also suggest that IMUs are a promising modality for detecting FoG episodes, as both ML and DL models built upon accelerometry data consistently performed well ( Table 1 ). Of note, the straightforward nature of sensor placement and removal would facilitate data collection and quality, allowing for greater accessibility across the PD patient population. While the vast majority of included studies report exclusively on predictive algorithms, detection is only the first step of FoG episode management. To effectively “close the loop” and improve patients’ quality of life, a therapeutic aspect ( Figure 7 ) must follow. Figure 7. Open in a new tab Example sensor locations on the body for ambulatory data collection during walking. White circles indicate representative anatomical sites where biomechanical and neural signals may be recorded, including the back, waist, thigh, and ankle. Pink highlighted circles represent intracranial DBS electrodes (local field potentials), a knee-mounted inertial measurement unit (IMU), and insole pressure sensors at the feet. Time-synchronized neural and biomechanical signals during walking and freezing-of-gait (FoG) episodes can be used to derive biomarkers that inform personalized, closed-loop therapeutic algorithms. 4.1. Closed-Loop Neuromodulation The closed-loop systems explored in the Results all provide intervention in the form of tactile or auditory stimulation to the user. As FoG is inherently a manifestation of a neuromechanical disorder, exploring this realm may confer more physiologically relevant therapy. In February 2025, Medtronic (Minneapolis, MN) received approval from the U.S. Food and Drug Administration to implement adaptive deep brain stimulation (aDBS) algorithms on the Percept™ PC device to treat motor symptoms in patients with Parkinson’s disease. The advent of sensing-and-stimulating, implantable hardware like Percept™ PC enables at-home electrophysiological data to be streamed and analyzed for biomarkers of pathological state, which can then be programmed back within the device. Unlike continuous DBS, which can exacerbate FoG by delivering constant stimulation irrespective of neural state, aDBS would dynamically adjust stimulation in real time based on specific neural signatures. In a landmark study, Oehrn et al. [ 80 ] demonstrate the long-term feasibility of aDBS in four patients by linking stimulation to local field potential (LFP)-derived control signals. To generate the biomarkers, the team applied neural recordings into a linear discriminant analysis classifier, a form of supervised learning. While FoG episodes were not specifically addressed by the study, the researchers tailored each patient’s neural biomarker to their individual motor symptom profile, such as bradykinesia or dyskinesia. Based on this framework, the unique neural signatures expressed during movement suggest considerable potential to personalize the algorithms used in aDBS to treat FoG. By synchronizing kinematic and LFP recordings from the Medtronic Summit RC+S device, one group identified pathological beta frequency (13–30 Hz) bursts in the STN as a candidate control signal for mitigating FoG in seven patients [ 81 ]. In addition to reducing FoG episode duration and frequency, the aDBS paradigm was shown to improve bradykinesia, walking tasks, and Unified Parkinson’s Disease Rating Scale (MDS-UPDRS III) scores. However, when the researchers compared outcomes with adaptive and traditional DBS, they found minimal differences, which they posit may be due to the need for a longer acclimation period for the newer paradigm. The extent to which a singular neural signature (e.g., beta bursts) can encapsulate underlying FoG pathophysiology is likely to fluctuate based on environmental factors [ 82 ]. Looking ahead, the discrepancies in clinical presentation (freezing while turning, start hesitation, etc.), the role of emotional processing, and the breadth of algorithmic control policies each warrant resolution when developing aDBS strategies for FoG [ 83 , 84 ]. Although all aDBS studies reported in this paper harnessed supervised learning techniques for biomarker discovery, future systems may incorporate a DL approach that automatically adjusts the neural control signal over time to account for temporal changes. However, adjusting stimulation settings with individual gait cycles means adapting at relatively short-term durations (i.e., milliseconds), when events occurring on longer timescales, like cognitive cues, may also confer an effect. Advances in reinforcement learning and brain state decoding could pave the way for algorithms that adapt across multiple timescales, from the FoG episode itself to more longitudinal measures like disease progression. For example, Horn et al. [ 85 ] propose the potential for aDBS algorithms to automatically switch between treating different PD symptom states, like tremor and rigidity, by decoding at the brain circuitry level. Termed the “adaptive circuit targeting” DBS framework, their system would incorporate connectome information to attain a more naturalistic form of therapy. Ultimately, a stronger mechanistic understanding of FoG will better inform how neuromodulation can be harnessed to manage or even prevent episodes. 4.2. Closed-Loop Biomechanics Despite algorithmic advances being important for accurate FoG prediction, the translation of these models into the real world as therapeutic tools depends on their biomechanical implementation. External devices such as robotic exoskeletons, soft robotic apparel, and insole-based systems have been explored as potential additions to ML-driven FoG detection [ 86 , 87 ]. These devices aim to provide mechanical cues or gait compensation during FoG episodes, thereby reducing fall risk. However, their effectiveness is limited by various biomechanical considerations. Any wearable device must maintain the body’s natural gait mechanics. Exoskeletons, for example, may stabilize joint motion but can also inadvertently alter arm swing, stride length, or add in compensatory movements, potentially creating maladaptive gait patterns over time [ 88 ]. Similarly, insole-based sensors show promise as a minimally invasive option, but prolonged use could affect comfort, weight distribution, and plantar pressure, thereby affecting gait [ 89 ]. In comparison, soft robotic garments offer improved flexibility and adaptability compared to rigid exoskeletons, but they still have to deliver precise torques to specific joints without interfering with natural movements, often incorporating cable actuators to augment hip flexion to combat FoG episodes [ 87 ]. Furthermore, patients may consciously alter the way they walk because they are aware of the device, or unconsciously compensate for added bulk, weight, or poorly positioned sensors [ 88 ]. These compensations could shift loading patterns across joints, muscles, and bones in ways that, over time, could lead to secondary musculoskeletal issues. For example, an insole that relieves FoG in the short term might subtly alter knee or hip biomechanics, increasing strain on proximal joints [ 90 ]. Likewise, heavier or inflexible devices could affect balance, posture, or arm swing [ 88 ]. These risks underscore the need for ergonomic, lightweight, and unobtrusive device designs, as well as longitudinal studies that evaluate both FoG reduction and longitudinal consequences on joint health and overall musculoskeletal function. Another challenge is in patient-specific biomechanics. Anatomical differences, disease states, and comorbidities influence how wearables interact with patients. For example, stride variability, muscle strength, and foot anatomy can alter the effectiveness of interventions [ 91 ]. Adaptive algorithms tailored to each patient’s biomechanics could help personalize therapy while minimizing unintended gait disruptions. This emphasizes the need for future directions in which wearable devices serve as both a diagnostic tool and as a device for individualized therapy and treatment that is based on the anatomy of each patient. Lastly, practical factors such as device comfort and integration into daily life are critical for implementing wearables in the real world. Real-world feasibility requires patients to tolerate these medical devices for long durations without interfering with ADLs. Translating prediction into effective closed-loop interventions will therefore require careful design of wearable hardware to ensure that biomechanical considerations do not undermine therapeutic benefit, and vice versa. 4.3. Translational Considerations Towards Real-Time Therapies Translating predictive architectures into meaningful therapies will require resolving several data-related challenges. While sophisticated algorithms may offer a more accurate prediction of upcoming FoG events, this trend creates a notable tradeoff between model complexity and robustness. An improvement in traditional metrics may not directly correlate with real-world application due to the technological constraints of wearable and implantable devices. Complex architectures incorporating elements like ensemble learning may require greater power consumption, making battery life a potential concern. In addition, the latency period between FoG detection and intervention should be within a certain range; otherwise, therapy (i.e., neurostimulation, torque forces, cueing) will not be delivered at clinically relevant timescales. For example, if aDBS algorithms modulate stimulation parameters at the gait-cycle level, the adjustment period of amplitude ramping may not align with the FoG episode, reducing efficacy. A limitation that spread across the included studies was relatively small participant cohorts. It is imperative that this limitation be taken into account when interpreting deep learning results. These models tend to perform better when they are trained on large datasets. Therefore, smaller sample sizes may increase the risk that performance reflects overfitting rather than true FoG features. Furthermore, the persistent reuse of the same public datasets, along with not testing enough on independent data, can inflate the reported accuracy. This diminishes confidence in how well these approaches may actually translate to real-world settings [ 92 ]. To better separate true FoG features from differences between patient groups, future studies should analyze results by medication status and task conditions (e.g., turning or dual-task walking), clearly report disease severity and duration, and use study designs or analysis approaches that account for differences between individuals. Because most of the studies used cohorts with 20 or fewer participants and because many papers utilized the same or similar datasets, e.g., Daphnet, the performance improvements reported may capture dataset-specific patterns rather than true freezing-of-gait physiology; this is especially the case with complex DL models. At a minimum, we suggest that future studies test the models on new patients, try to use independent datasets, and lastly, report practical performance details (the number of detected episodes, response timing, and energy use, for example) in conjunction with the standard accuracy scores to avoid reporting potentially overly optimistic conclusions. Towards addressing these concerns, the rise in the “Internet of Things” across healthcare and various industries has ushered in a new wave of improvements in data processing efficiency. As explored in Section 3.5 , intentional algorithmic design choices can greatly enhance the translational potential of models. Dvorani et al. [ 19 ] demonstrate that edge computing can reduce latency and maximize computational resources; delegating calculations to a distributed network of sensors may thus be a valuable strategy for deploying closed-loop systems, especially in biomechanical solutions such as exoskeletons. Moreover, while the DL models in this study tended to provide greater performance than their traditional ML counterparts, they also require considerably more data for training. Ongoing advances in personalized algorithm design can help optimize real-world scalability and also address the patient-specific variability of FoG. Transfer learning is one computational approach that could balance the need for individualized models with the aforementioned constraints. The method involves first training a foundational AI model to solve a given task (e.g., FoG management in Patient #1), which can later be reapplied to solve related tasks (e.g., Patient #2). Transfer learning operates on the principle that there is a generalizable basis across tasks or patients, such that initially learned features are applicable to novel scenarios. Rather than expend extensive resources to collect FoG episode data and build a model from scratch for each patient, pre-trained transformer models can be finetuned with smaller datasets [ 93 , 94 ], reducing computational, patient, and healthcare provider burden. Finally, taking into consideration prior insights on IMUs ( Table 2 ), optimizing the location of fewer total sensors can reduce model complexity, while still retaining high performance. In particular, O’Day et al. found that on a FoG prediction task, a single ankle IMU had similar performance (AUC = 0.80) to the best combination of IMUs in the study, which was the lower back and both ankles (AUC = 0.83) [ 39 ]. We see similar examples of this trend in the included studies: Chan et al. found that a CNN trained on a single IMU placed at the waist had 83% accuracy [ 17 ]. Meanwhile, Yang et al. trained a temporal CNN on data from five IMUs on the pelvis and lower body, which conferred an F1-score of 0.63 [ 30 ]. However, direct comparison across studies remains challenging due to differences in patient populations, model architectures, and the nature of reported metrics. Future research studies focused on IMU optimization and patient preferences can improve the real-world translation of data-driven insights. 4.4. Limitations This review highlights several limitations in the current landscape of machine learning applications for FoG management. First, there was substantial heterogeneity in reported performance metrics. Whereas some studies emphasized only the accuracy or F1-score, others focused on sensitivity, specificity, or AUC, making direct comparisons across models challenging. Moreover, due to the varying influence of class imbalance on each metric, the one that a given paper may have optimized its model on could directly affect its translational potential. With this in mind, standardized reporting frameworks could better facilitate proper benchmarking and reproducibility. Of note, the closed-loop papers discussed in Section 3.5 did not utilize the number of falls as an outcome measure, despite its notable relevance to the goals of FoG reduction. Implementing more clinically meaningful metrics like fall reduction, real-time power consumption, and patient comfort will be important considerations for validating future closed-loop devices. Second, the size and diversity of study cohorts were limited. A majority of the studies had fewer than 20 patients, and many drew from the same publicly available datasets, such as the Daphnet Freezing of Gait dataset. Small cohort sizes were consistently a limitation across the studies included in this review. This is particularly problematic with the use of deep learning. Deep learning models often require large and diverse datasets to establish stable and generalizable representations. Thus, training on limited samples, repeated use of publicly available datasets, and limited testing on independent external datasets may contribute to high performance estimates and diminished generalizability to the real world [ 92 ]. These constraints are further influenced by clinical heterogeneity both within and across study populations. Variabilities in age, disease duration and severity, and medication status are known to influence the occurrence and presentation of FoG, yet these factors are inconsistently accounted for, which limits the model interpretability and translational relevance. Having small and homogeneous samples limits the generalizability of findings, specifically when one considers the variability in FoG presentation across disease states, demographics, and comorbidities. While each included paper reported on the patient cohort size, the vast majority did not provide statistics regarding the total count of individual labeled FoG episodes, which would provide more granular information on dataset quality. In addition, the motor impairment severity of patients was rarely reported, and usually not in a standardized manner, i.e., MDS-UPDRS III scores. Improving data collection to include larger and more diverse populations could help ensure that algorithms maximize external validity to real-world settings. When the number was reported, FoG episodes were often relatively scarce compared to the total duration of gait recordings, leading to overrepresentation of normal walking patterns. The usage of data augmentation strategies like synthetic oversampling and generative AI in the resulting studies warrants further evaluation, stressing the need for more nuanced metrics to evaluate model performance. Models built with these features may present with high accuracy in a computational setting, but may not take into account contextual and behavioral factors associated with mobility. While synthetic oversampling methods like SMOTE were applied in some cases, these approaches cannot entirely substitute for high-quality, naturally balanced data. Without substantive external validation, it remains uncertain whether these approaches to handling imbalanced biomedical data truly augment performance. More extensive long-term monitoring could help capture a better distribution of FoG events. Additionally, the data collection environment was underreported and varied considerably across studies, with most recordings conducted in controlled clinical settings where FoG was induced by standardized tasks (e.g., TUG). While these environments ensure consistent labeling, they may not reflect the complexity of FoG in day-to-day life. The reliance on induced FoG using standardized clinical tests, such as the Timed Up and Go and dual task test, raises important questions about the ecological validity of the reported findings and generalization to at-home settings or nursing homes [ 95 , 96 ]. Ambulatory, at-home monitoring is needed to capture more naturalistic episodes and environments, which will be critical for building algorithms that are adaptable to real-life scenarios and true to each patient’s needs and space. These limitations highlight the necessity for more comprehensive, real-world datasets and standardized evaluation practices. Addressing these gaps will be essential for advancing ML- and DL-based FoG detection into reliable, clinically useful tools. Finally, although this review emphasized real-time detection and closed-loop therapeutic applications, only a small minority of the included studies implemented a true “closed-loop” structure. Key practical considerations for real-world employment (detection latency, computational burden, clinically meaningful outcomes, and fall reduction) were inconsistently reported. This observation showcases a gap between the current translational goals in a clinical setting and the available evidence in the literature. As such, our statements regarding real-time and closed-loop feasibility should be interpreted as hypothesis-generating rather than confirmatory. In addition to the limitations within the reviewed literature, this review itself also comes with its own limitations. Firstly, the search strategy employed in this review focused on studies that referenced “freezing of gait,” specifically. This confined search strategy may have led to the exclusion of relevant works that included FoG characteristics without explicitly using that terminology. This restriction may have influenced the scope and representativeness of the included literature and should be considered when interpreting the findings of this review. 5. Conclusions As FoG is a major contributor to the number of falls that a patient with PD may experience—and falls are closely tied to increased risk of injury and hospitalization—it is critical to establish accurate detection and prediction systems that improve patient safety. This review highlights both the progress and the challenges that ML and DL methods provide within sensor-based FoG management. These high-throughput approaches can be applied to identify FoG episode biomarkers and generate therapies in a personalized manner. In particular, IMUs and algorithms such as CNNs and LSTMs consistently demonstrated strong performance, emphasizing the potential for an objective gait monitoring framework based on these combined techniques. However, this review also shows that translation into real-world practice and the use of closed-loop therapies remain limited. As only a small fraction of the included studies utilize a closed-loop framework, practical challenges to implementation may not yet have been uncovered or resolved. Obstacles in contemporary research include the use of small, homogeneous datasets, inconsistent reporting standards, and the need to integrate biomechanical and neurophysiological aspects. Future work should prioritize the adoption of more diverse, naturalistic datasets and distributed computing strategies that optimize therapeutic parameters in real time, as well as clinically relevant measures like fall reduction and latency. Looking forward, physiology-driven devices that leverage ML could offer valuable benefits for patients with PD, including reduced FoG, reduced fall risk, and improved mobility and quality of life. Acknowledgments The authors consent that all figures were created by the authors. Abbreviations The following abbreviations are used in this manuscript: PD Parkinson’s disease FoG Freezing of gait ADL Activities of daily living IMU Inertial measurement unit CNN Convolutional neural network LSTM Long short-term memory SMOTE Synthetic minority oversampling ML Machine learning DL Deep learning kNN K nearest neighbors SVM Support vector machine GenAI Generative Artificial Intelligence AI Artificial intelligence ECoG Electrocorticography DBS Deep brain stimulation TUG Timed Up and Go RNN Recurrent neural network LFP Local field potential EMG Electromyography EEG Electroencephalography STN Subthalamic nucleus AUC Area Under the Curve Open in a new tab Supplementary Materials The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/s26072042/s1 . sensors-26-02042-s001.zip (138.5KB, zip) Author Contributions Conceptualization, M.S., M.B., M.E.H. and S.A.H.; methodology, M.S., M.B. and S.A.H.; formal analysis, M.S.; investigation, M.S. and C.G.; writing—original draft preparation, M.S. and C.G.; writing—review and editing, M.B., M.E.H. and S.A.H.; visualization, M.S.; supervision, M.B., M.E.H. and S.A.H.; All authors have read and agreed to the published version of the manuscript. Institutional Review Board Statement Not applicable. Data Availability Statement No new data were created or analyzed in this study. Conflicts of Interest The authors declare no conflicts of interest. Funding Statement This research received no external funding. Footnotes Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. References 1. Freezing Parkinson’s Foundation. [(accessed on 8 February 2025)]. Available online: https://www.parkinson.org/living-with-parkinsons/management/activities-daily-living/freezing . 2. Ge H.L., Chen X.Y., Lin Y.X., Ge T.J., Yu L.H., Lin Z.Y., Wu X.Y., Kang D.Z., Ding C.Y. The prevalence of freezing of gait in Parkinson’s disease and in patients with different disease durations and severities. Chin. Neurosurg. J. 2020;6:17. doi: 10.1186/s41016-020-00197-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Nonnekes J., Bloem B.R. Biphasic Levodopa-Induced Freezing of Gait in Parkinson’s Disease. J. Park. Dis. 2020;10:1245–1248. doi: 10.3233/JPD-201997. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Gizem T., des Neiges Santin M., Olivier C., Collomb-Clerc A., Fernandez-Vidal S., Hainque E., Bardinet E., Lau B., François C., Karachi C., et al. Freezing of gait depends on cortico-subthalamic network recruitment following STN-DBS in PD patients. Park. Relat. Disord. 2022;104:49–57. doi: 10.1016/j.parkreldis.2022.10.002. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Tosserams A., Fasano A., Gilat M., Factor S.A., Giladi N., Lewis S.J.G., Moreau C., Bloem B.R., Nieuwboer A., Nonnekes J. Management of Freezing of Gait—Mechanism-Based Practical Recommendations. Nat. Rev. Neurol. 2025;21:327–344. doi: 10.1038/s41582-025-01079-6. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Boukhennoufa I., Zhai X., Utti V., Jackson J., McDonald-Maier K.D. Wearable sensors and machine learning in post-stroke rehabilitation assessment: A systematic review. Biomed. Signal Process. Control. 2022;71:103197. doi: 10.1016/j.bspc.2021.103197. [ DOI ] [ Google Scholar ] 7. Sengupta N., Rao A.S., Yan B., Palaniswami M. A survey of wearable sensors and machine learning algorithms for Automated Stroke Rehabilitation. IEEE Access. 2024;12:36026–36054. doi: 10.1109/ACCESS.2024.3373910. [ DOI ] [ Google Scholar ] 8. Ricotti V., Kadirvelu B., Selby V., Festenstein R., Mercuri E., Voit T., Faisal A.A. Wearable full-body motion tracking of activities of daily living predicts disease trajectory in Duchenne muscular dystrophy. Nat. Med. 2023;29:95–103. doi: 10.1038/s41591-022-02045-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Al-Tashi Q., Saad M.B., Muneer A., Qureshi A., Mirjalili S., Sheshadri A., Le X., Vokes N.I., Zhang J., Wu J. Machine learning models for the identification of prognostic and Predictive cancer biomarkers: A systematic review. Int. J. Mol. Sci. 2023;24:7781. doi: 10.3390/ijms24097781. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Palmqvist S., Tideman P., Mattsson-Carlgren N., Schindler S.E., Smith R., Ossenkoppele R., Calling S., West T., Monane M., Verghese P.B., et al. Blood Biomarkers to Detect Alzheimer Disease in Primary Care and Secondary Care. JAMA. 2024;332:1245–1257. doi: 10.1001/jama.2024.13855. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Slieker R.C., Donnelly L.A., Akalestou E., Lopez-Noriega L., Melhem R., Güneş A., Azar F.A., Efanov A., Georgiadou E., Muniangi-Muhitu H., et al. Identification of biomarkers for glycaemic deterioration in type 2 diabetes. Nat. Commun. 2023;14:2533. doi: 10.1038/s41467-023-38148-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Joshi G., Jain A., Araveeti S.R., Adhikari S., Garg H., Bhandari M. FDA-Approved Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices: An Updated Landscape. Electronics. 2024;13:498. doi: 10.3390/electronics13030498. [ DOI ] [ Google Scholar ] 13. Salomon A., Gazit E., Ginis P., Urazalinov B., Takoi H., Yamaguchi T., Goda S., Lander D., Lacombe J., Sinha A.K., et al. A machine learning contest enhances automated freezing of gait detection and reveals time-of-day effects. Nat. Commun. 2024;15:4853. doi: 10.1038/s41467-024-49027-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Tricco A.C., Lillie E., Zarin W., O’Brien K.K., Colquhoun H., Levac D., Moher D., Peters M.D.J., Horsley T., Weeks L., et al. PRISMA Extension for Scoping Reviews (PRISMA-ScR): Checklist and explanation. Ann. Intern. Med. 2018;169:467–473. doi: 10.7326/M18-0850. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Slemenšek J., Geršak J., Bratina B., van Midden V.M., Pirtošek Z., Šafarič R. Wearable Online Freezing of Gait Detection and Cueing System. Bioengineering. 2024;11:1048. doi: 10.3390/bioengineering11101048. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Koltermann K., Clapham J., Blackwell G., Jung W., Burnet E.N., Gao Y., Shao H., Cloud L., Pretzer-Aboff I., Zhou G. Proceedings of the 2024 IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE) Volume 2024. IEEE; New York, NY, USA: 2024. Gait-Guard: Turn-aware Freezing of Gait Detection for Non-intrusive Intervention Systems; pp. 61–72. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Chan L.L.Y., Yang S., Aswani M., Kark L., Henderson E., Lord S.R., Brodie M.A. Development, Validation, and Limits of Freezing of Gait Detection Using a Single Waist-Worn Device. IEEE Trans. Biomed. Eng. 2024;71:3024–3031. doi: 10.1109/TBME.2024.3407059. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Al-Nefaie A.H., Aldhyani T.H.H., Farhah N., Koundal D. Intelligent diagnosis system based on artificial intelligence models for predicting freezing of gait in Parkinson’s disease. Front. Med. 2024;11:1418684. doi: 10.3389/fmed.2024.1418684. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Dvorani A., Wiesener C., Salchow-Hommen C., Jochner M., Spieker L., Skrobot M., Voigt H., Kühn A., Wenger N., Schauer T. On-Demand Gait-Synchronous Electrical Cueing in Parkinson’s Disease using Machine Learning and Edge Computing-A Pilot Study. IEEE Open J. Eng. Med. Biol. 2024;5:306–315. doi: 10.1109/OJEMB.2024.3390562. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Li W., Chen X., Zhang J., Lu J., Zhang C., Bai H., Liang J., Wang J., Du H., Xue G., et al. Recognition of Freezing of Gait in Parkinson’s Disease Based on Machine Vision. Front. Aging Neurosci. 2022;14:921081. doi: 10.3389/fnagi.2022.921081. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Mesin L., Porcu P., Russu D., Farina G., Borzì L., Zhang W., Guo Y., Olmo G. A multi-modal analysis of the freezing of gait phenomenon in parkinson’s disease. Sensors. 2022;22:2613. doi: 10.3390/s22072613. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Liu D.F., Zhao B.T., Zhu G.Y., Liu Y.Y., Bai Y.T., Liu H.G., Jiang Y., Zhang X., Shi L., Zhang H., et al. Synchronized Intracranial Electrical Activity and Gait Recording in Parkinson’s Disease Patients with Freezing of Gait. Front. Neurosci. 2022;16:795417. doi: 10.3389/fnins.2022.795417. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Borzì L., Mazzetta I., Zampogna A., Suppa A., Olmo G., Irrera F. Prediction of Freezing of Gait in Parkinson’s Disease Using Wearables and Machine Learning. Sensors. 2021;21:614. doi: 10.3390/s21020614. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Pardoel S., Shalin G., Nantel J., Lemaire E.D., Kofman J. Early Detection of Freezing of Gait during Walking Using Inertial Measurement Unit and Plantar Pressure Distribution Data. Sensors. 2021;21:2246. doi: 10.3390/s21062246. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Zhang Y., Yan W., Yao Y., Ahmed J.B., Tan Y., Gu D. Prediction of Freezing of Gait in Patients with Parkinson’s Disease by Identifying Impaired Gait Patterns. IEEE Trans. Neural Syst. Rehabil. Eng. 2020;28:591–600. doi: 10.1109/TNSRE.2020.2969649. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Demrozi F., Bacchin R., Tamburin S., Cristani M., Pravadelli G. Toward a Wearable System for Predicting Freezing of Gait in People Affected by Parkinson’s Disease. IEEE J. Biomed. Health Inform. 2020;24:2444–2451. doi: 10.1109/JBHI.2019.2952618. [ DOI ] [ PubMed ] [ Google Scholar ] 27. Abbasi S., Rezaee K. Deep Learning–Based Prediction of Freezing of Gait in Parkinson’s Disease with the Ensemble Channel Selection Approach. Brain Behav. 2024;15:e70206. doi: 10.1002/brb3.70206. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Wang B., Hu X., Ge R., Xu C., Zhang J., Gao Z., Zhao S., Polat K. Prediction of Freezing of Gait in Parkinson’s disease based on multi-channel time-series neural network. Artif. Intell. Med. 2024;154:102932. doi: 10.1016/j.artmed.2024.102932. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Sun R., Hu K., Ehgoetz Martens K.A., Hagenbuchner M., Tsoi A.C., Bennamoun M., Lewis S.J.G., Wang Z. Higher Order Polynomial Transformer for Fine-Grained Freezing of Gait Detection. IEEE Trans. Neural Netw. Learn. Syst. 2024;35:12746–12759. doi: 10.1109/TNNLS.2023.3264647. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Yang P.K., Filtjens B., Ginis P., Goris M., Nieuwboer A., Gilat M., Slaets P., Vanrumste B. Freezing of gait assessment with inertial measurement units and deep learning: Effect of tasks, medication states, and stops. J. Neuroeng. Rehabil. 2024;21:24. doi: 10.1186/s12984-024-01320-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Park J.M., Moon C.W., Lee B.C., Oh E., Lee J., Jang W.J., Cho K.H., Lee S.H. Detection of freezing of gait in Parkinson’s disease from foot-pressure sensing insoles using a temporal convolutional neural network. Front. Aging Neurosci. 2024;16:1437707. doi: 10.3389/fnagi.2024.1437707. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Klaver E.C., Heijink I.B., Silvestri G., van Vugt J.P.P., Janssen S., Nonnekes J., van Wezel R.J.A., Tjepkema-Cloostermans M.C. Comparison of state-of-the-art deep learning architectures for detection of freezing of gait in Parkinson’s disease. Front. Neurol. 2023;14:1306129. doi: 10.3389/fneur.2023.1306129. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Borzì L., Sigcha L., Olmo G. Context Recognition Algorithms for Energy-Efficient Freezing-of-Gait Detection in Parkinson’s Disease. Sensors. 2023;23:4426. doi: 10.3390/s23094426. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Hu K., Mei S., Wang W., Ehgoetz Martens K.A., Wang L., Lewis S.J.G., Feng D.D., Wang Z. Multi-Level Adversarial Spatio-Temporal Learning for Footstep Pressure Based FoG Detection. IEEE J. Biomed. Health Inform. 2023;27:4166–4177. doi: 10.1109/JBHI.2023.3272902. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Hou Y., Ji J., Zhu Y., Dell T., Liu X. Flexible Gel-Free Multi-Modal Wireless Sensors with Edge Deep Learning for Detecting and Alerting Freezing of Gait Symptom. IEEE Trans. Biomed. Circuits Syst. 2023;17:1010–1021. doi: 10.1109/TBCAS.2023.3281596. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Borzì L., Sigcha L., Rodriguez-Martin D., Olmo G. Real-time detection of freezing of gait in Parkinson’s disease using multi-head convolutional neural networks and a single inertial sensor. Artif. Intell. Med. 2023;135:102459. doi: 10.1016/j.artmed.2022.102459. [ DOI ] [ PubMed ] [ Google Scholar ] 37. Kwon H., Clifford G.D., Genias I., Bernhard D., Esper C.D., Factor S.A., McKay J.L. An Explainable Spatial-Temporal Graphical Convolutional Network to Score Freezing of Gait in Parkinsonian Patients. Sensors. 2023;23:1766. doi: 10.3390/s23041766. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Shi B., Tay A., Au W.L., Tan D.M.L., Chia N.S.Y., Yen S.C. Detection of Freezing of Gait Using Convolutional Neural Networks and Data From Lower Limb Motion Sensors. IEEE Trans. Biomed. Eng. 2022;69:2256–2267. doi: 10.1109/TBME.2022.3140258. [ DOI ] [ PubMed ] [ Google Scholar ] 39. O’Day J., Lee M., Seagers K., Hoffman S., Jih-Schiff A., Kidziński L., Delp S., Bronte-Stewart H. Assessing inertial measurement unit locations for freezing of gait detection and patient preference. J. Neuroeng. Rehabil. 2022;19:20. doi: 10.1186/s12984-022-00992-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Filtjens B., Ginis P., Nieuwboer A., Afzal M.R., Spildooren J., Vanrumste B., Slaets P. Modelling and identification of characteristic kinematic features preceding freezing of gait with convolutional neural networks and layer-wise relevance propagation. BMC Med. Inform. Decis. Mak. 2021;21:341. doi: 10.1186/s12911-021-01699-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Shi B., Yen S.C., Tay A., Tan D.M.L., Chia N.S.Y., Au W.L. Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020. IEEE; New York, NY, USA: 2020. Convolutional Neural Network for Freezing of Gait Detection Leveraging the Continuous Wavelet Transform on Lower Extremities Wearable Sensors Data; pp. 5410–5415. [ DOI ] [ PubMed ] [ Google Scholar ] 42. Pardoel S., Nantel J., Kofman J., Lemaire E.D. Prediction of Freezing of Gait in Parkinson’s Disease Using Unilateral and Bilateral Plantar-Pressure Data. Front. Neurol. 2022;13:831063. doi: 10.3389/fneur.2022.831063. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Bikias T., Iakovakis D., Hadjidimitriou S., Charisis V., Hadjileontiadis L.J. DeepFoG: An IMU-Based Detection of Freezing of Gait Episodes in Parkinson’s Disease Patients via Deep Learning. Front. Robot. AI. 2021;8:537384. doi: 10.3389/frobt.2021.537384. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Mostafa T.A., Soltaninejad S., McIsaac T.L., Cheng I. A Comparative Study of Time Frequency Representation Techniques for Freeze of Gait Detection and Prediction. Sensors. 2021;21:6446. doi: 10.3390/s21196446. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Shalin G., Pardoel S., Lemaire E.D., Nantel J., Kofman J. Prediction and detection of freezing of gait in Parkinson’s disease from plantar pressure data using long short-term memory neural-networks. J. Neuroeng. Rehabil. 2021;18:167. doi: 10.1186/s12984-021-00958-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Dvorani A., Waldheim V., Jochner M.C.E., Salchow-Hömmen C., Meyer-Ohle J., Kühn A.A., Wenger N., Schauer T. Real-Time Detection of Freezing Motions in Parkinson’s Patients for Adaptive Gait Phase Synchronous Cueing. Front. Neurol. 2021;12:720516. doi: 10.3389/fneur.2021.720516. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Miao R., Shokur S., De Lima Pardini A.C., Boari D., Bouri M. Proceedings of the 2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA) IEEE; New York, NY, USA: 2021. Detecting Freezing of Gait in Parkinson’s Disease Patient via Deep Residual Network; pp. 320–325. [ DOI ] [ Google Scholar ] 48. Pardoel S., Shalin G., Lemaire E.D., Kofman J., Nantel J. Grouping successive freezing of gait episodes has neutral to detrimental effect on freeze detection and prediction in Parkinson’s disease. PLoS ONE. 2021;16:e0258544. doi: 10.1371/journal.pone.0258544. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Sigcha L., Costa N., Pavón I., Costa S., Arezes P., López J.M., De Arcas G. Deep Learning Approaches for Detecting Freezing of Gait in Parkinson’s Disease Patients through On-Body Acceleration Sensors. Sensors. 2020;20:1895. doi: 10.3390/s20071895. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Li B., Yao Z., Wang J., Wang S., Yang X., Sun Y. Improved Deep Learning Technique to Detect Freezing of Gait in Parkinson’s Disease Based on Wearable Sensors. Electronics. 2020;9:1919. doi: 10.3390/electronics9111919. [ DOI ] [ Google Scholar ] 51. Liu X., Chen S., Ren K., Zhao J. Proceedings of the 2020 Chinese Control and Decision Conference (CCDC) IEEE; New York, NY, USA: 2020. A Freezing of Gait Detection System Based on Stepwise Feature Extraction Method; pp. 2060–2065. [ DOI ] [ Google Scholar ] 52. Hu K., Wang Z., Wang W., Ehgoetz Martens K., Wang L., Tan T., Lewis S., Feng D. Graph Sequence Recurrent Neural Network for Vision-Based Freezing of Gait Detection. IEEE Trans. Image Process. 2020;29:1890–1901. doi: 10.1109/TIP.2019.2946469. [ DOI ] [ PubMed ] [ Google Scholar ] 53. Borzi L., Olmo G., Artusi C.A., Lopiano L. Proceedings of the 2022 IEEE 46th Annual Computers, Software, and Applications Conference (COMPSAC) IEEE; New York, NY, USA: 2020. Detection of Freezing of Gait in People with Parkinson’s Disease using Smartphones; pp. 625–635. [ DOI ] [ Google Scholar ] 54. Reches T., Dagan M., Herman T., Gazi E., Gouskova N., Giladi N., Manor B., Hausdorff J. Using wearable sensors and machine learning to automatically detect freezing of gait during a fog-provoking test. Sensors. 2020;20:4474. doi: 10.3390/s20164474. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Shalin G., Pardoel S., Nantel J., Lemaire E.D., Kofman J. Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) IEEE; New York, NY, USA: 2020. Prediction of freezing of gait in parkinson’s disease from foot plantar-pressure arrays using a convolutional neural network; pp. 244–247. [ DOI ] [ PubMed ] [ Google Scholar ] 56. Filtjens B., Nieuwboer A., D’cruz N., Spildooren J., Slaets P., Vanrumste B. A data-driven approach for detecting gait events during turning in people with Parkinson’s disease and freezing of gait. Gait Posture. 2020;80:130–136. doi: 10.1016/j.gaitpost.2020.05.026. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Martinez-Villasenor L., Ponce H., Nunez-Martinez J. Proceedings of the 2020 International Conference on Mechatronics, Electronics and Automotive Engineering (ICMEAE) IEEE; New York, NY, USA: 2020. Artificial Hydrocarbon Networks for Freezing of Gait Detection in Parkinso’s Disease; pp. 3–8. [ DOI ] [ Google Scholar ] 58. Unni M.P., Menon P.P., Livi L., Wilson M., Young W., Bronte-Stewart H., Tsaneva-Atanasova K. Data-Driven Prediction of Freezing of Gait Events From Stepping Data. Front. Med. Technol. 2020;2:581264. doi: 10.3389/fmedt.2020.581264. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Ezhilarasi J., Senthil Kumar T. Develop a novel, faster mask region-based convolutional neural network model with leave-one-subject-out to predict freezing of gait abnormalities of parkinson’s disease. Neural Comput. Appl. 2025;37:5441–5457. doi: 10.1007/s00521-024-10832-9. [ DOI ] [ Google Scholar ] 60. Kumar P.D., Sowmya B.J., Vincent A., Pandiyan A. Proceedings of the 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT) IEEE; New York, NY, USA: 2024. Attention based 1-D CNN system for detection of Freezing of gait (FOG) in Parkinsons Patients; pp. 1–7. [ DOI ] [ Google Scholar ] 61. Kondo Y., Bando K., Suzuki I., Miyazaki Y., Nishida D., Hara T., Kadone H., Suzuki K. Video-based detection of freezing of gait in daily clinical practice in patients with parkinsonism. IEEE Trans. Neural Syst. Rehabil. Eng. 2024;32:2250–2260. doi: 10.1109/TNSRE.2024.3413055. [ DOI ] [ PubMed ] [ Google Scholar ] 62. Pardoel S., AlAkhras A., Jafari E., Kofman J., Lemaire E.D., Nantel J. Real-Time Freezing of Gait Prediction and Detection in Parkinson’s Disease. Sensors. 2024;24:8211. doi: 10.3390/s24248211. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Yang P.K., Filtjens B., Ginis P., Goris M., Nieuwboer A., Gilat M., Slaets P., Vanrumste B. Automatic Detection and Assessment of Freezing of Gait Manifestations. IEEE Trans. Neural Syst. Rehabil. Eng. 2024;32:2699–2708. doi: 10.1109/TNSRE.2024.3431208. [ DOI ] [ PubMed ] [ Google Scholar ] 64. Banu A.N.S. Proceedings of the 2024 First International Conference for Women in Computing (InCoWoCo) IEEE; New York, NY, USA: 2024. Enhancing Fog Detection with Wavelet Transforms using Random Forest and Neural Networks; pp. 1–8. [ DOI ] [ Google Scholar ] 65. Bacanin N., Petrovic A., Jovanovic L., Zivkovic M., Zivkovic T., Sarac M. Proceedings of the 2024 5th International Conference on Mobile Computing and Sustainable Informatics (ICMCSI) IEEE; New York, NY, USA: 2024. Parkinson’s Disease Induced Gain Freezing Detection using Gated Recurrent Units Optimized by Modified Crayfish Optimization Algorithm. [ DOI ] [ Google Scholar ] 66. Peppes N., Tsakanikas P., Daskalakis E., Alexakis T., Adamopoulou E., Demestichas K. FoGGAN: Generating Realistic Parkinson’s Disease Freezing of Gait Data Using GANs. Sensors. 2023;23:8158. doi: 10.3390/s23198158. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Dimoudis D., Tsolakis N., Magga-Nteve C., Meditskos G., Vrochidis S., Kompatsiaris I. InSEption: A Robust Mechanism for Predicting FoG Episodes in PD Patients. Electronics. 2023;12:2088. doi: 10.3390/electronics12092088. [ DOI ] [ Google Scholar ] 68. Li B., Li Y., Sun Y., Yang X., Zhou X., Yao Z. A monitoring method of freezing of gait based on multimodal fusion. Biomed. Signal Process. Control. 2023;82:104589. doi: 10.1016/j.bspc.2023.104589. [ DOI ] [ Google Scholar ] 69. Önder M., Şentürk Ü., Polat K., Paulraj D. Proceedings of the 2023 International Conference on Research Methodologies in Knowledge Management, Artificial Intelligence and Telecommunication Engineering (RMKMATE) IEEE; New York, NY, USA: 2023. Classification of freezing of gait in parkinson’s disease using machine learning algorithms; pp. 1–5. [ DOI ] [ Google Scholar ] 70. Koltermann K., Jung W., Blackwell G., Pinney A., Chen M., Cloud L., Pretzer-Aboff I., Zhou G. Proceedings of the IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies. Volume 2023. Association for Computing Machinery; New York, NY, USA: 2023. FoG-Finder: Real-time Freezing of Gait Detection and Treatment; pp. 22–33. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 71. Sadiq M., Tauheed Khan M., Masood S. Attention-Based Deep Learning Model for Early Detection of Parkinson’s Disease. Comput. Mater. Contin. 2022;71:5183–5200. doi: 10.32604/cmc.2022.020531. [ DOI ] [ Google Scholar ] 72. Mekruksavanich S., Jantawong P., Jitpattanakul A. Proceedings of the 2022 Joint International Conference on Digital Arts, Media and Technology with ECTI Northern Section Conference on Electrical, Electronics, Computer and Telecommunications Engineering (ECTI DAMT & NCON) IEEE; New York, NY, USA: 2022. Comparative Analysis of CNN-based Deep Learning Approaches on Complex Activity Recognition; pp. 338–341. [ DOI ] [ Google Scholar ] 73. Ren K., Chen Z., Ling Y., Zhao J. Recognition of freezing of gait in Parkinson’s disease based on combined wearable sensors. BMC Neurol. 2022;22:229. doi: 10.1186/s12883-022-02732-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Filtjens B., Ginis P., Nieuwboer A., Slaets P., Vanrumste B. Automated freezing of gait assessment with marker-based motion capture and multi-stage spatial-temporal graph convolutional neural networks. J. Neuroeng. Rehabil. 2022;19:48. doi: 10.1186/s12984-022-01025-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. Manna S.K., Hannan Bin Azhar M.A., Greace A. Optimal locations and computational frameworks of FSR and IMU sensors for measuring gait abnormalities. Heliyon. 2023;9:e15210. doi: 10.1016/j.heliyon.2023.e15210. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Sun Y., Song Z., Mo L., Li B., Liang F., Yin M., Wang D. IMU-Based quantitative assessment of stroke from gait. Sci. Rep. 2025;15:9541. doi: 10.1038/s41598-025-94167-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Roggen D., Plotnik M., Hausdorff J. Daphnet Freezing of Gait [Dataset] UCI Machine Learning Repository; Irvine, CA, USA: 2010. [ DOI ] [ Google Scholar ] 78. Chawla N.V., Bowyer K.W., Hall L.O., Kegelmeyer W.P. SMOTE: Synthetic Minority Over-sampling Technique. J. Artif. Intell. Res. 2002;16:321–357. doi: 10.1613/jair.953. [ DOI ] [ Google Scholar ] 79. Hu J., Shen L., Albanie S., Sun G., Wu E. Squeeze-and-Excitation Networks. arXiv. 2019 doi: 10.1109/TPAMI.2019.2913372.1709.01507 [ DOI ] [ PubMed ] [ Google Scholar ] 80. Oehrn C.R., Cernera S., Hammer L.H., Shcherbakova M., Yao J., Hahn A., Wang S., Ostrem J., Little S., Starr P. Chronic adaptive deep brain stimulation versus conventional stimulation in parkinson’s disease: A blinded randomized feasibility trial. Nat. Med. 2024;30:3345–3356. doi: 10.1038/s41591-024-03196-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 81. Wilkins K.B., Petrucci M.N., Lambert E.F., Melbourne J., Gala A., Akella P., Parisi L., Cui C., Kehnemouyi Y., Hoffman S. Beta burst-driven adaptive deep brain stimulation for gait impairment and freezing of gait in parkinson’s disease. Brain Commun. 2025;7:fcaf266. doi: 10.1093/braincomms/fcaf266. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 82. Oliveira R., Coelho L., Rebouças E., Ferreira-Pinto M.J., Vaz R., Aguiar P. Machine learning for adaptive deep brain stimulation in Parkinson’s disease: Closing the loop. J. Neurol. 2023;270:5313–5326. doi: 10.1007/s00415-023-11873-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 83. Klocke P., Loeffler M.A., Lewis S.J.G., Gharabaghi A., Weiss D. Could adaptive deep brain stimulation treat freezing of gait in Parkinson’s disease? J. Neurol. 2025;272:267. doi: 10.1007/s00415-025-13000-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 84. Neumann W.J., Gilron R., Little S., Tinkhauser G. Adaptive Deep Brain Stimulation: From Experimental Evidence Toward Practical Implementation. Mov. Disord. 2023;38:937–948. doi: 10.1002/mds.29415. [ DOI ] [ PubMed ] [ Google Scholar ] 85. Horn A., Neumann W.J. From adaptive deep brain stimulation to adaptive circuit targeting. Nat. Rev. Neurol. 2025;21:556–566. doi: 10.1038/s41582-025-01131-5. [ DOI ] [ PubMed ] [ Google Scholar ] 86. Marcante A., Di Marco R., Gentile G., Pellicano C., Assogn F., Pontieri F., Spalletta G., Macchiusi L., Gatsios D., Giannakis A., et al. Foot Pressure Wearable Sensors for Freezing of Gait Detection in Parkinson’s Disease. Sensors. 2021;21:128. doi: 10.3390/s21010128. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 87. Kim J., Porciuncula F., Yang H.D., Wendel N., Baker T., Chin A., Ellis T., Walsh C. Soft robotic apparel to avert freezing of gait in Parkinson’s disease. Nat. Med. 2024;30:177–185. doi: 10.1038/s41591-023-02731-8. [ DOI ] [ PubMed ] [ Google Scholar ] 88. Swank C., Wang-Price S., Gao F., Almutairi S. Walking with a Robotic Exoskeleton Does Not Mimic Natural Gait: A Within-Subjects Study. JMIR Rehabil. Assist. Technol. 2019;6:e11023. doi: 10.2196/11023. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 89. Tramonti C., Griguoli F., Martino S.D., Lamola G., Falossi F., Palattella G., Raffaetà G., Chisari C., Parchi P. The Effect of Proprioceptive Insoles on Gait Kinematics in Healthy Participants. JPO J. Prosthet. Orthot. 2024;36:273–280. doi: 10.1097/JPO.0000000000000500. [ DOI ] [ Google Scholar ] 90. Lack S., Barton C., Woledge R., Laupheimer M., Morrissey D. The immediate effects of foot orthoses on hip and knee kinematics and muscle activity during a functional step-up task in individuals with patellofemoral pain. Clin. Biomech. 2014;29:1056–1062. doi: 10.1016/j.clinbiomech.2014.08.005. [ DOI ] [ PubMed ] [ Google Scholar ] 91. Jalloul N. Wearable sensors for the monitoring of movement disorders. Biomed. J. 2018;41:249–253. doi: 10.1016/j.bj.2018.06.003. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 92. Vabalas A., Gowen E., Poliakoff E., Casson A.J. Machine learning algorithm validation with a limited sample size. PLoS ONE. 2019;14:e0224365. doi: 10.1371/journal.pone.0224365. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 93. Chato L., Regentova E. Survey of Transfer Learning Approaches in the Machine Learning of Digital Health Sensing Data. J. Pers. Med. 2023;13:1703. doi: 10.3390/jpm13121703. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 94. Sharma A., Lysenko A., Jia S., Boroevich K.A., Tsunoda T. Advances in AI and machine learning for predictive medicine. J. Hum. Genet. 2024;69:487–497. doi: 10.1038/s10038-024-01231-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 95. Del Din S., Godfrey A., Mazzà C., Lord S., Rochester L. Free-living monitoring of Parkinson’s disease: Lessons from the field. Mov. Disord. 2016;31:1293–1313. doi: 10.1002/mds.26718. [ DOI ] [ PubMed ] [ Google Scholar ] 96. Mancini M., Weiss A., Herman T., Hausdorff J.M. Turn Around Freezing: Community-Living Turning Behavior in People with Parkinson’s Disease. Front. Neurol. 2018;9:18. doi: 10.3389/fneur.2018.00018. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials sensors-26-02042-s001.zip (138.5KB, zip) Data Availability Statement No new data were created or analyzed in this study. Articles from Sensors (Basel, Switzerland) are provided here courtesy of Multidisciplinary Digital Publishing Institute (MDPI) ACTIONS View on publisher site PDF (1.6 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Freezing of Gait in Parkinson's Disease: A Heterogeneous Phenomenon Across Assessment Tools#273302
• Use of CereGate Therapy for Freezing of Gait in PD#877610
• Randomized Safety Study of CVT-301 Compared to an Observational Control Group#1172333
• Long-Term Safety, Tolerability and Efficacy in Perampanel Treated Parkinson's Disease Patients With Motor Fluctuations#1101350
• Cellular and Molecular Aspects of Parkinson Treatment: Future Therapeutic Perspectives#935861