RecallRisk-BERT: A Multi-Task Framework for Post-Report Medical Device Recall Triage Ali Semih Atalaya , Sevgi Yigit-Serta
arXiv:2606.27174v1 [cs.LG] 25 Jun 2026
a
Department of Computer Engineering, Faculty of Engineering, Ankara University, Ankara, 06830, Türkiye
Abstract Medical device recalls are a critical regulatory mechanism for protecting patient safety. The growing volume of FDA recall records presents challenges in post-report recall triage, severity assessment, and root-cause interpretation. Existing studies mostly address recall occurrence prediction or root-cause analysis separately, while joint modeling of recall severity and root-cause categories has received limited attention. We develop an automated recall triage framework using 54,165 FDA medical device recall records from openFDA, covering the period from 2002 to October 2025. We first evaluate classical machine learning and boosting-based models for recall severity and rootcause category prediction. We then develop RecallRisk-BERT, a multi-task model that combines PubMedBERT-based textual representations of recall narratives with embedding-based representations of structured categorical features, including product code, regulation number, and medical specialty. The model simultaneously predicts recall severity (Class I/II/III) and a consolidated root-cause category (9 classes). Performance was evaluated using accuracy, macro-averaged precision, recall, F1-score, and ROC-AUC. In single-task severity prediction, our LightGBM-based text–tabular configuration achieved the strongest performance, with an accuracy of 0.963, macro-F1 of 0.856, and ROC-AUC of 0.974. In the multi-task setting, RecallRiskBERT substantially outperformed the single-task PubMedBERT baseline. Model-derived risk rankings were strongly consistent with observed rootcause severity patterns (ρ = 0.983, p = 1.936 × 10−6 ). These findings indicate that text–tabular learning can support scalable post-report recall triage, regulatory decision support, and model-based root-cause risk analysis. Keywords: Medical device recall, Multi-task learning, Text–tabular fusion, PubMedBERT, Recall triage, FDA
1. Introduction Medical devices have become an integral and vital part of healthcare services. With rapid advancements in technology and science, medical devices play an important role in modern medicine and make significant contributions to healthcare delivery. Especially after the COVID-19 pandemic, the role of these devices in healthcare provision has gained strategic importance [11, 13]. They improve quality of life across a wide spectrum, ranging from routine healthcare tools to life-saving and life-changing devices. Medical devices encompass a broad variety of products, from simple items such as medical gloves and bandages to complex machines like blood pressure monitors and X-ray devices. The development and production of medical devices is a complex and demanding process that requires high-quality standards and ensures that the final product complies with existing regulatory requirements. Moreover, maintaining quality throughout this process must be meticulously documented [16]. This process has a multi-stage and interdisciplinary structure, extending from conceptual design to a fully functional product ready for clinical use [2]. At any of these stages—such as design, manufacturing, or software development—deviations in quality and safety may lead to medical device recall cases. Medical devices that comply with standards and are available on the market are continuously monitored, and any emerging issues are promptly recalled. A recall is the process of removing or correcting products that violate standards and is overseen by the the U.S. Food and Drug Administration (FDA). Recalls are voluntary actions taken by manufacturers and distributors to protect public health and welfare by withdrawing products that pose significant risks, such as injury, or are otherwise defective. The protocol known as 21 CFR Part 7 provides guidance to responsible companies for conducting effective recalls. However, in some cases, if a manufacturer or importer does not voluntarily recall a device that poses a health risk, the FDA can issue a mandatory recall order under 21 CFR Part 810 (Medical Device Recall Authority). Hundreds of medical device recalls occur each year [6]. It has been reported that, in 2022, FDA recorded 70 high-risk recalls, compared to the fiveyear average of 47 [12]. According to Sarkissian [11], “packaging” (47.4%), “component” (14%) and “design” (13.3%) were the dominant reasons in a significant portion of high-risk recalls. Medical device recalls lead to sig2
nificant financial losses for manufacturers [8], reputational damage [3], and disruptions in the supply chain [1, 7]. At the same time, they pose direct clinical risks to patients and healthcare professionals. Therefore, regulatory processes are of critical importance for patient safety and the sustainability of healthcare systems. Devices regulated by the FDA are classified into three categories based on the level of risk they pose to patients: Class I (low risk), Class II (moderate risk), and Class III (high risk). Medical device regulatory classes and recall severity classes use different risk logics. While Class III devices generally indicate higher regulatory device risk, Class I recalls involve situations where exposure to a violative product may reasonably cause serious health problems or death. Class II recalls involve products that may cause temporary or medically reversible adverse health consequences, and Class III recalls involve products unlikely to cause adverse health consequences [5]. Life-supporting devices, such as implantable cardioverter defibrillators, are categorized as Class III devices; however, when a recall of such devices poses the highest level of patient risk, it is classified as a Class I recall [11, 14]. Since this classification directly reflects the potential clinical impact of a recall, it is critically important not only to model whether a recall will occur but also to accurately model the severity of the recall. The increasing prevalence of recalls has triggered growing interest in proactive prediction models within the literature. A significant portion of the studies on medical device recalls has focused on binary classification problems aimed at predicting whether a recall will occur or not. More recent studies have proposed approaches for predicting recall causes (root-cause or recall initiator) as multi-class classification problems. However, existing studies generally address these two problems independently and do not jointly model the relationship between recall severity and the underlying error mechanism. Yet, it is known that different error mechanisms produce different levels of clinical risk. The FDA’s recall classification process itself evaluates factors such as the nature and potential health risk of the harm, the scope and cause of the defect, the probability of occurrence, and the level of danger posed to patients together [5]. Empirical evidence also confirms this structural relationship. Software design-related failures account for approximately 42% of AI/ML-enabled medical device recalls [25]. A study on Class I recalls shows that design defects, software-related failures, and sterility issues are among the major root-causes associated with high-risk recalls, while labeling and packaging problems are also among the frequently reported recall causes [26]. This situation indicates that there is a structural connection 3
between recall severity and root-cause. To the best of our knowledge, an integrated (joint) prediction framework that models these two targets together is still absent in the literature. Existing approaches generally consider either severity class or root-cause alone as a single target, which prevents the models from benefiting from cross-task information sharing. Moreover, the structural characteristics of FDA recall databases create significant class imbalance problems from a modeling perspective. While Class I recalls occur less frequently compared to Class II recalls, certain root-cause categories — especially software design in AI/ML devices — appear to be highly concentrated [25, 36]. Such imbalanced distributions may create bias toward majority classes, particularly in single-task classification models, and may make it more difficult to distinguish minority classes. The accurate identification of high-risk recalls, which are especially critical for patient safety, makes this problem even more important. Within the scope of this study, the following research questions were investigated: • RQ1: To what extent can recall severity be reliably predicted using recall narratives and device/regulatory context? • RQ2: Does jointly learning recall severity and root-cause category (i.e., recall cause) improve prediction performance compared to single-task models? • RQ3: Which root-cause categories are associated with higher risk severity? To answer these research questions, a multi-task learning framework that jointly learns recall severity (recall class) and recall cause (root-cause category) is proposed. In the proposed approach, recall texts (reason_for_recall and product_description) are represented using a transformer-based model pretrained on biomedical texts (e.g., PubMedBERT). In addition, these features are combined with structured (tabular) data such as product code, regulation number, and medical specialty area to construct a rich feature space. Over this shared representation, two different tasks are learned simultaneously: (i) prediction of recall severity and (ii) classification of recall cause. Thus, the model aims not only to improve prediction performance, but also to generate mechanistic and explainable insights regarding recall processes. 4
The performance of the proposed approach was evaluated comparatively against classical machine learning methods (Random Forest, XGBoost, LightGBM), text-based hybrid models (e.g., SentenceBERT/BioBERT embeddings + boosting), and single-task transformer-based models. In addition, by analyzing the risk representations learned by the model, the effects of different error mechanisms on recall severity were examined through both data-driven (ground-truth) and model-based analyses. In this context, the study aims not only to propose a prediction model, but also to provide datadriven and interpretable insights into medical device safety. The main contributions of this study can be summarized as follows: • The predictability of recall severity is systematically investigated using a broad set of machine learning and deep learning models, • A multi-task learning framework, RecallRisk-BERT, jointly modeling recall severity and root-cause category is proposed. • A text–tabular fusion architecture combining text-based representations and structured data is developed. • The relationship between different error mechanisms and recall severity is investigated through both statistical (ground-truth) and model-based analyses. • The proposed model is comprehensively compared with strong singletask baselines, including classical machine learning, boosting, and transformer-based approaches. The aim of this study is not to predict whether a future medical device will be recalled, since all observations in the dataset correspond to recall events that have already occurred. Rather, this work addresses the problem of post-report recall triage: given an issued recall report, the proposed model jointly predicts the severity level and identifies the underlying root-cause mechanism in a unified framework. This formulation directly responds to the operational needs of regulatory authorities and manufacturers, who must rapidly assess and prioritize reported recalls under resource constraints. The proposed framework can therefore be interpreted as an intelligent decisionsupport tool for prioritizing issued recall reports according to severity and root-cause risk. 5
The rest of the paper is organized as follows: Section 2 reviews the literature on artificial intelligence in medical devices; Section 3 describes the dataset and methodology; Section 4 explains Experimental Setup, and Section 5 presents the experimental results. The final section discusses the implications and limitations of the study. 2. Related Work The studies on medical devices have expanded significantly in both scope and methodological diversity over the past fifteen years. This section first summarizes the development of artificial intelligence (AI) in medical devices, then reviews the literature on prediction and analysis of medical device recalls. The last section compiles the key gaps in the existing literature and justifies which gaps this study focuses on. 2.1. AI in Medical Devices Artificial intelligence (AI) and machine learning (ML) technologies in health services and biomedical engineering have led to important developments in recent years. AI-based methods have been used for various health and research purposes, including Disease diagnosis, treatment planning, chronic illness management, medicine discovery, and clinical decision support systems. In line with these developments, medical device manufacturers have also begun integrating AI and machine learning technologies into their devices to improve product performance, enhance diagnostic accuracy, and improve patient outcomes [18, 19, 20, 21]. The rapid increase in the number of AI-powered medical devices has also attracted the attention of regulatory bodies. According to reports published by the FDA, the number of AI-based medical devices, mostly in the field of radiology [22], is increasing every year, indicating the growing adoption of AI in healthcare technologies. As of October 2023, the number of AI and machine learning-based devices approved by the FDA was reported as 691. Of these, 531 (77%) were in radiology, 70 (10%) in cardiovascular surgery, 20 in neurology, 15 in hematology, and 55 in other fields [23]. Furthermore, it is reported that the emergence of Large Language Models (LLMs) has accelerated the development processes of AI-powered medical devices and created new application areas [24]. With the proliferation of AI-based devices, their safety profiles have become a separate area of research. Since 2019, the FDA has intensified its work 6
on policies regarding the authorization of these devices. Chen et al. [25] analyzed AI/ML-powered medical device recalls over a 27-year data window from 1997–2024 and found that software design flaws explained a significant portion of recalls in this category. This finding suggests that AI-based devices have different risk profiles than traditional medical devices and that conventional recall analysis frameworks may be insufficient. Therefore, analyzing and predicting medical device failures and recalls has become a critical research area for both manufacturers and regulatory bodies. 2.2. Prediction and analysis of medical device recalls Medical device recalls are one of the most important regulatory mechanisms to protect patient safety. Regulatory bodies such as the FDA classify recalls into three categories based on the health risk they pose: Class I (highest risk), Class II (medium risk), and Class III (lowest risk). Therefore, the analysis of recall records has emerged as an important research topic for assessing health risks and improving product safety. Descriptive and Epidemiological Studies. Villarraga et al. [14] established the basic conceptual framework of the field by presenting a descriptive analysis of FDA recalls. Thirumalai and Sinha [13] empirically examined the sources of recalls in the medical device industry and their financial consequences, reporting that device design, manufacturing processes, and labeling errors were the primary triggers. More recently, Sarkissian [11] analyzed 871 Class I recalls from 2014–2018, reporting that the dominant root-causes were “packaging” (47.4%), “component” (14%), and “design” (13.3%). M.J. et al. [26] examined Class I recalls from 2020–2023 and reported that injury risks, erroneous results, software problems, and manufacturing defects are the dominant root-cause categories. While these studies offer valuable descriptive findings, they have remained largely descriptive and have not provided opportunities for proactive prediction. Recall Prediction and Classification Approaches. The use of machine learning methods on medical device recall data has become increasingly common in recent years. Emakhu et al. [49] developed a machine learning framework to predict recall failure types using software-based medical device recall records published by the FDA; approximately 84% accuracy was achieved with the multilayer perceptron (MLP) model. Slivinskis et al. [2] presented a random forest-based algorithm using data from open sources such as PubMed and Google Trends; they demonstrated that they could predict recall status with high accuracy 3, 6, and 12 months in advance. The strength of the study 7
is that it provides evidence of the predictive power of signals from external (non-FDA) data sources; its limitation is its small sample size and the use of external media/literature signals rather than internal device features. Zhu et al. [29] proposed a deep learning-based prediction structure on a dataset covering 45,398 devices validated via 510(k) paths between 2003 and 2020. Their method models the structural features of the predicate citation network using graph learning methods and captures the temporal patterns of citation network features. The results showed that the network-based approach provided significant improvements in both accuracy and timeliness compared to classical tabular methods. Everhart et al. [27], published in JAMA, found statistically significant relationships between the predictive properties of 510(k) devices and subsequent recalls using linear probability models. Zhalechian et al. [28] developed a gradient-boosting-based human–algorithm framework to estimate future recall risk for more than 31,000 FDA 510(k) devices and support clearance decisions under FDA workload constraints. Root-Cause Estimation Approaches. Hu et al. [30] employed a machine learning–natural language processing workflow to analyze medical device recall causes, using DBSCAN clustering to identify recall initiator categories and text-similarity-based classification to support practitioner-oriented insights. In their later work, Hu et al. [48] argued that the medical device recall problem should be addressed from a supply chain risk management and resilience perspective and developed a recall initiator prediction framework that combines optimized feature selection and clustering techniques to strengthen error detection at the premarket stage. Highlighting the contribution of practical features such as “product code” to predictive power, the study achieved an accuracy level of 88.85% for high-risk Class II and Class III devices. Table 1 provides a comparative overview of the methodologies of existing studies. 2.3. Research Gap This study is designed to explicitly address the following gaps in the medical device recall estimation literature. Gap 1 — The severity level is not directly modeled as a multi-class target. The vast majority of studies either treat the recall event as a binary (recall/norecall) target or focus only on Class I recalls. Zhalechian et al. [28] optimized their models for a binary target, then reported individual AUC values posthoc according to severity classes; however, their models were not structurally designed for a multi-class severity target. A systematic estimation framework 8
Table 1: Comparison of methodological characteristics of major studies in the medical device recall prediction literature. Study
Prediction Target
Task Type
Sarkissian [11]
Root-cause dis- Descriptive tribution
Emakhu et al. [49]
Failure type
Everhart et al. [27]
Recall rence
Methodology
Input ties
Statistical analysis
Tabular only
Multiclass clas- MLP + ML sification
occur- Binary tion
Predic- Linear probability
Hu et al. [30]
Recall initiator
MJ et al. [26]
Root-cause dis- Descriptive tribution
Unsupervised
Modali-
Tabular only Tabular only
DBSCAN + NLP
Text only
Statistical analysis
Tabular only
Zhalechian et al. Recall [28] rence
occur- Binary classifi- Gradient boosting cation
Tabular only
Slivinskis et al. [2]
Recall rence
occur- Binary classifi- Random forest cation
Tabular + external signals (PubMed/Trends)
Chen et al. [25]
Root-cause dis- Descriptive tribution
Statistical analysis
Hu et al. [48]
Recall initiator
Random forest + Tabular only DBSCAN
Zhu et al. [29]
Recall rence
This study
Recall sever- Multi-task ity + root- multiclass cause
Multiclass
Tabular only
occur- Binary classifi- Graph + temporal Tabular + netcation DL work structure Text + tabular PubMedBERT + tabular fusion (multimodal) + MTL
Note: DL: Deep Learning; MTL: Multi-Task Learning; MLP: Multi-Layer Perceptron; DBSCAN: Density-Based Spatial Clustering of Applications with Noise.
that sets Class I/Class II/Class III severity levels as a multi-class classification problem from the outset and is specifically designed for this problem is not common in the literature. Gap 2 — The severity and root-cause dimensions are treated separately. Current approaches either target only the recall event or only the rootcause; a multi-tasking framework where these two outputs are learned together through shared representations has not been found in the literature. 9
However, for regulatory decision support systems, both pieces of information—how dangerous the device is and why it is dangerous—must be accessible simultaneously. To the best of our knowledge, the multi-task learning (MTL) paradigm has not yet been adapted to the context of medical device recall prediction. This study addresses this gap by proposing an MTL framework that simultaneously predicts severity and root-cause through shared representations. Gap 3 — Insufficient integration of textual and tabular data. Although FDA recall records contain rich textual descriptions (reason for recall, product descriptions) and regulatory tabular fields (product code, regulation number, specialty area), current prediction studies often treat these two data types separately: some use only tabular features, while others model only textual information. Although the success of biomedical domain-specific language models (PubMedBERT [51], BioBERT [52], ClinicalBERT [56]) in health text classification has been demonstrated in various studies [53, 54], the application of these models to direct recall severity estimation and their use within a multimodal architecture with tabular features has not been systematically addressed in the literature. This study responds to this gap by adopting a multimodal learning approach that combines biomedical BERT representations with tabular feature embeddings. 3. Materials and Methods 3.1. Data source In this study, medical device recall records from the FDA covering the period from November 2002 to October 2025 were used. The dataset consists of 54,165 recall records obtained in JSON format from the openFDA platform1 . The study included all medical device categories available in the FDA medical device recall database, without imposing any restriction in terms of medical specialty area or approval pathway. From the raw openFDA records, 11 relevant fields (as shown in Table 2) were retained for dataset construction, target definition, preprocessing, and model development. The primary inputs of the model consisted of two textual fields (reason for recall and product description) and three structured categorical fields (product code, regulation number, and medical specialty description). The remaining fields were used 1
https://open.fda.gov/apis/device/recall/download/
10
(a)
(b)
Figure 1: Class distributions of (a) recall class and (b) root-cause category. The numeric labels correspond to Class I, Class II, and Class III recall categories.
for target construction, descriptive analysis, or sensitivity experiments, as described in the following sections. In particular, recall class was used as the primary target variable, while root-cause description was used only to derive the consolidated root-cause category target and was not included as an input feature. The dataset has a significant class imbalance in both recall severity and root-cause dimensions. In the recall class variable, Class II constitutes the dominant majority with 87.5%, while Class I is represented at 7.9% and Class III at 4.6%. The distribution of the secondary target variable, root-cause category, is: Unknown (28.8%), Design (27.1%), Manufacturing (14.0%), Material (13.4%), Packaging (6.6%), Labeling (4.6%), Human (2.5%), Regulatory (2.2%), and Software (0.7%). The distribution of target classes is shown in Figure 1. 3.2. Target variable construction The study defines two principal target variables: recall class and rootcause category. The former one is defined by the FDA and categorizes recall events into three classes: Class I, Class II, and Class III. Class I recalls represent the highest-risk recalls, carrying a probability of serious health risk or death; Class II recalls represent moderate-risk recalls that may lead to temporary or medically reversible health problems; and Class III recalls represent the lowest-risk recalls, which are not expected to pose a serious health risk. The secondary target variable is the root-cause category variable, which expresses the fundamental underlying cause mechanism of the recall. It is derived from the root-cause description field contained in the FDA records. 11
Table 2: Selected openFDA fields used for dataset construction and modeling. Field product res number product description product code
recalling firm
reason for recall root-cause description device name medical specialty description regulation number recall class event date initiated
Description Internal recall identifier for each recalled device. Brief description of the recalled product. Three-letter code identifying the device category, based on its 21 CFR classification, technology, and intended use. Firm initiating the recall or primarily responsible for the product’s manufacture/marketing. Description of the product defect and how it violates the FD&C Act. FDA-determined general type of recall cause.
Type Identifier String Categorical
Categorical
String Categorical
Proprietary or trade name of the cleared device. Medical specialty assigned by the device regulation.
Categorical
CFR regulation under which the device is classified. Recall classification based on the reason for recall. Date the firm first notified the public or consignees.
Categorical
Categorical
Integer Date
Rather than using the raw root-cause descriptions directly, similar causes were consolidated under more general categories. This restructuring aims to enable the model to learn more reliable root-cause groups and to alleviate the learning difficulty posed by rare or ambiguous causes. Rare or ambiguous root-causes were grouped under the “Unknown” category. The final root-cause category target variable consists of 9 main categories: Unknown, Design, Manufacturing, Material, Packaging, Labeling, Human, Regulatory, and Software. To obtain a continuous model-derived risk estimate, the predicted class probabilities from the recall severity head were converted into an Expected Recall Risk Score. Since FDA Class I recalls represent the highest risk level and Class III recalls represent the lowest risk level, the class probabilities 12
were weighted as follows: Expected Recall Risk Score = 3P (Class I) + 2P (Class II) + 1P (Class III)
(1)
This score ranges from 1 to 3, where higher values indicate a higher modelestimated recall risk severity. The score was used only for model-based risk analysis and was not used as an additional training target. 3.3. Feature construction and leakage control Model inputs were divided into two groups: textual features and structured features. The textual input was constructed by concatenating the reason for recall and product description fields using a separator token and subsequently tokenized with a maximum sequence length of 128 tokens to improve computational efficiency. The structured inputs consisted of the product code, regulation number, and medical specialty description fields. Each categorical variable was mapped to an integer identifier, and an “UNK” token was assigned to categories not observed in the training set. The target variables were converted into numerical representations using label encoding. To mitigate potential data leakage, the device name and recalling firm variables were excluded from the input feature set. These variables may introduce shortcut learning by enabling the model to exploit firm-specific recall patterns or device-name-specific associations rather than learning generalizable relationships from the underlying recall characteristics. Note that the root-cause description field was not used as an input feature, because it was used to derive the secondary target variable, root-cause category. 3.4. Proposed RecallRisk-BERT architecture This study proposes a multi-task hybrid fusion model based on BERT, called RecallRisk-BERT, for post-report medical device recall triage. Figure 2 illustrates the hybrid architecture and decision-support workflow of RecallRisk-BERT. The framework first processes recall narratives and structured device/regulatory fields, then jointly predicts recall severity and rootcause category through a shared multi-task representation. The predicted severity probabilities are subsequently transformed into an Expected Recall Risk Score to support post-report recall triage. 13
The model consists of four components: (i) a transformer-based text encoder, (ii) a tabular embedding module, (iii) shared representation layers, and (iv) task-specific output heads. Text encoder. The combined text input is encoded with PubMedBERT [51], which has been documented to exhibit superior performance in biomedical terminology. The domain-specific terminology density of FDA recall records, such as manufacturing defects, sterility breaches, contamination, and software problems, constitutes the main rationale for this architectural choice. The contextual vector of the [CLS] token is used as the text representation. Tabular embedding module. Three categorical variables are passed through separate embedding layers and transformed into learnable dense vectors. After concatenation, they are passed through an MLP layer to obtain a 128dimensional tabular representation. Shared representation and task-specific heads. A 768-dimensional [CLS] output of PubMedBERT is combined with a tabular representation to form an 896-dimensional fused vector; this vector is then passed through a fully linked layer and transformed into a 512-dimensional shared representation. Two parallel classification heads produce logit outputs for recall class (3 classes) and root-cause category (9 classes), respectively. 3.5. Multi-task learning objective The proposed model is based on a hard parameter-sharing approach: the transformer-based text encoder, the tabular fusion component, and the shared representation layers are shared by both tasks; only the final output heads are separated in a task-specific manner. This design enables the structural dependencies between recall severity and root-cause mechanisms to be implicitly learned within a shared latent representation. The total loss function is defined as a linear combination of task-specific weighted cross-entropy losses: Ltotal = λ1 Lrecall_severity + λ2 Lroot_cause_category (2) To address class imbalance, inverse-frequency-based class weights were scaled according to Equation 3 to prevent the instability caused by excessively large values. N (3) wi = K · ni where N is the total number of samples, K is the number of classes, and ni is the number of samples for class i. 14
Figure 2: Proposed RecallRisk-BERT architecture integrating PubMedBERT-based textual representations and tabular embeddings for joint prediction of recall severity and root-cause category.
15
4. Experimental Setup Baseline Models: To evaluate the effectiveness of the proposed RecallRiskBERT model, single-task baseline models from different model families were employed. The first group consists of classical machine learning models trained with TF-IDF-based text representations. In this group, logistic regression and Support Vector Machine (SVM) [38] were used as fundamental text classification baselines. The second group consisted of boosting-based models, including Random Forest [39], XGBoost [40], LightGBM [41]. The third group consists of single-task transformer-based models. These models employed a BERT-based text encoder and a tabular fusion structure, but were trained to predict only a single target variable. In this way, a single-task severity prediction model, a single-task root-cause category prediction model, and the proposed multi-task RecallRisk-BERT model were compared. This comparison was used to assess whether the performance difference originates from shared representation learning. In addition to the classical machine learning models, two neural baseline architectures were implemented to evaluate the effectiveness of nontransformer deep learning approaches, using GloVe-based [58] text representations combined with tabular features. The first baseline is a deep neural network (DNN) consisting of four fully connected layers with 1024, 512, 256, and 128 neurons, respectively, each followed by batch normalization and dropout to mitigate overfitting, with ReLU activation in the hidden layers and softmax in the output layer for multi-class classification. The second baseline is a BiLSTM model [59] with an attention mechanism. Here, the GloVe-based textual features are processed by a bidirectional LSTM that captures contextual dependencies from both past and future directions, while the attention layer assigns higher weights to the most informative tokens of the recall narrative. The resulting text representation is then concatenated with the tabular features and passed through three fully connected layers of 512, 256, and 128 neurons, each followed by batch normalization and dropout, with a softmax output layer. For both the DNN and BiLSTM baselines, three task settings are evaluated: single-task recall severity prediction (Y1), single-task root-cause category prediction (Y2), and multi-task prediction of Y1 and Y2. Training protocol: We adopted a stratified random 80%–20% train–test split based on the primary target variable, recall class, to preserve the recall severity distribution. For classical ML and boosting-based models, 1016
fold cross-validation was performed on the training set, and the validation trends were found to be consistent with the final test results. For DNN and BiLSTM+Attention models, a 70%–10%–20% train–validation–test split was used due to the computational cost of repeated cross-validation. The validation set was used for early stopping and learning-rate reduction through ReduceLROnPlateau. For PubMedBERT-based models, the 80%–20% train– test split was used due to the high computational cost of transformer training. All experiments were conducted using a fixed random seed to ensure reproducibility of the train–test split and model initialization. Since the reported results are based on a single stratified train–test split, evaluating model robustness over multiple random seeds or repeated stratified splits is left for future work. We additionally experimented with a chronological (time-based) split using the recall initiation date, which resulted in a substantial performance drop, likely due to temporal distribution shift in recall reasons and device types over the study period; these results are omitted here for brevity. Since the objective of this study is to model the relationship between recall descriptions and severity/root-cause categories rather than to forecast future recalls, the stratified random split was selected as the primary evaluation protocol. Evaluation metrics: Model performance was evaluated using overall accuracy, macro-averaged precision, macro-averaged recall, and macro-averaged F1-score, and macro-averaged one-vs-rest ROC-AUC. Parameters: DNN and BiLSTM+Attention models were trained using the Adam optimizer with a learning rate of 5×10−4 and a batch size of 64. These models were trained for a maximum of 50 epochs, with early stopping based on validation loss. For PubMedBERT-based models, the Adam optimizer was used with a learning rate of 2 × 10−5 , a batch size of 128, and a maximum of 10 epochs. PubMedBERT was used as a frozen feature extractor; therefore, only the tabular embedding, fusion, shared representation, and task-specific output layers were trained. Model selection for PubMedBERTbased configurations was performed through staged experiments due to the high computational cost of transformer training. Class-weighted cross-entropy was used for the recall severity task to account for the imbalanced class distribution. The class weights used in the loss function are presented in Table 3. Since the raw weight value for the Software class was excessively high, it was additionally scaled by half and set to 8.0. In this main multi-task setting, equal task-level coefficients were 17
Table 3: Class weights computed for recall class and root-cause category.
Target variable recall class
root-cause category
Class Class I Class II Class III Unknown Design Manufacturing Material Packaging Labeling Human Regulatory Software
Weight 4.20 0.38 7.30 0.39 0.41 0.79 0.83 1.67 2.42 4.42 4.97 8.00
used, with (λ1 = λ2 = 1), assigning equal importance to recall severity prediction and root-cause category prediction. This configuration is reported in the main model comparison tables as RecallRisk-BERT. As an additional sensitivity experiment, an alternative empirically weighted multi-task loss configuration, RecallRisk-BERT-W, was also evaluated. The recall severity class weights were set to [3.5, 0.5, 6.0], and the root-cause category weights were set to [0.5, 0.5, 0.9, 1.0, 1.8, 2.5, 3.5, 3.8, 5.5]. Since these weights were empirically specified rather than selected through exhaustive hyperparameter optimization, this configuration is reported as a sensitivity analysis rather than as the primary model setting. Additional experiments with unequal task-weight configurations (i.e., λ1 andλ2 ) were also conducted; however, these settings did not improve the performance. 5. Results 5.1. Recall Severity Prediction Performance Table 4 presents a comparative evaluation of classical machine learning, deep learning, transformer-based, and multi-task learning approaches for recall severity prediction. Among the evaluated models, LightGBM with GloVe and tabular features achieved the highest accuracy, precision, F1-score, and ROC-AUC, indicating the strong predictive capability of boosting-based 18
Table 4: Performance comparison of machine learning, deep learning, transformer-based, and multi-task models for recall severity prediction. Recall, precision, and F1-score are reported as macro-averaged metrics to account for class imbalance. The best value for each metric is shown in bold. Model
Accuracy
Recall
Precision
F1
ROC-AUC
0.895
0.841
0.678
0.739
0.948
0.945
0.822
0.811
0.816
0.937
LightGBM
0.963
0.794
0.956
0.856
0.974
Boosting
XGBoost
0.960
0.793
0.938
0.850
0.967
Deep learning Deep learning
BiLSTM + Attention
0.933
0.854
0.772
0.807
0.954
0.928
0.833
0.756
0.791
0.948
Text + Tabular
GloVe + Tabular
TF-IDF + Tabular
Input Classical ML Classical ML
Logistic Regression Linear SVM
Boosting
DNN
Transformer
PubMedBERT
0.902
0.847
0.732
0.768
0.954
Proposed
RecallRisk-BERT
0.951
0.849
0.833
0.841
0.958
models for single-task recall severity classification. The highest macro recall was obtained by the BiLSTM+Attention model, although the margin was small, with RecallRisk-BERT also achieving a comparable macro recall of 0.849. The proposed RecallRisk-BERT model achieved competitive Y1 performance, with an accuracy of 0.951, a macro recall of 0.849, and an F1-score of 0.841. Although LightGBM provided the strongest overall single-task performance, RecallRisk-BERT substantially outperformed the single-task PubMedBERT baseline, whose F1-score was 0.768. Since both models use the same text and tabular inputs, this improvement cannot be attributed to the structured features alone; rather, it indicates that jointly learning recall severity together with root-cause category provides a clear benefit over an otherwise comparable single-task transformer baseline. Therefore, the role of RecallRisk-BERT is not to replace all classical machine learning approaches, but to provide a complementary multi-task framework that jointly models recall severity and root-cause mechanisms. Overall, the results show that both boosting-based machine learning models and transformer-based multi-task models are effective for post-report recall triage. While LightGBM provides the strongest single-task predictive performance for recall severity classification, RecallRisk-BERT offers an additional advantage by linking severity prediction with root-cause modeling 19
Text + Tabular
GloVe+ Tabular
Table 5: Performance comparison of deep learning, transformer-based, and multi-task models for recall severity and root-cause prediction. Y1 denotes recall severity prediction, while Y2 denotes root-cause category prediction. F1-score and recall are reported as macro-averaged metrics. RecallRisk-BERT-W denotes the empirically softened weighted multi-task configuration.
Model BiLSTM + Attention BiLSTM + Attention DNN DNN
Task setting
Y1 - F1
Y1 - Recall
Y2 - F1
Single-task Y1
0.807
0.854
—
Single-task Y2
—
—
0.657
Single-task Y1 Single-task Y2
0.791 —
0.833 —
— 0.661
BiLSTM + Attention
Multi Task (Y1 + Y2)
0.754
0.851
0.688
DNN
Multi Task (Y1 + Y2)
0.763
0.842
0.671
PubMedBERT
Single-task Y1
0.768
0.847
—
PubMedBERT
Single-task Y2
—
—
0.744
0.841
0.849
0.720
0.843
0.863
0.749
RecallRisk-BERT RecallRisk-BERT-W
Multi Task (Y1 + Y2) Multi Task (Y1 + Y2)
and model-based risk analysis. 5.2. Deep Learning and Multi-task Model Comparison Table 5 compares recurrent deep learning, feed-forward neural, transformerbased, and multi-task models for recall severity prediction (Y1) and rootcause category prediction (Y2). For Y1, among the standard model configurations, RecallRisk-BERT achieved the strongest overall multi-task performance by jointly predicting both recall severity and root-cause category within a single shared representation. Compared with the single-task PubMedBERT baseline, RecallRisk-BERT substantially improved Y1 macro-F1 from 0.768 to 0.841 while maintaining a comparable Y1 macro recall value of 0.849. In addition, unlike single-task models, RecallRisk-BERT simultaneously produced root-cause category predictions, achieving a Y2 F1-score of 0.720. 20
When compared with non-transformer neural baselines, RecallRisk-BERT also outperformed both the multi-task BiLSTM+Attention model and the multi-task DNN model in terms of Y1 F1-score and Y2 F1-score. This indicates that the combination of PubMedBERT-based contextual text representations and shared multi-task learning provides a more effective framework than recurrent or feed-forward neural architectures using GloVe-based representations. RecallRisk-BERT-W represents an empirically adjusted weighting variant of the proposed model, which applies a softened set of recall severity and rootcause category weights. Within the RecallRisk-BERT family, RecallRiskBERT-W further improved performance over the standard RecallRisk-BERT model, increasing Y1 F1-score from 0.841 to 0.843, Y1 macro recall from 0.849 to 0.863, and Y2 F1-score from 0.720 to 0.749. These results suggest that smoothing excessively large class weights can enhance both recall severity sensitivity and root-cause category prediction performance. Overall, these results indicate that while boosting-based models remain highly competitive for single-task recall severity classification, multi-task models provide a unified framework for jointly modeling recall severity and root-cause mechanisms, supporting broader functional coverage and more interpretable post-report recall triage. RecallRisk-BERT-W further shows that empirically softened class-weighting can improve the balance between severity sensitivity and root-cause prediction performance in the multi-task setting. 5.3. Effect of Text Representation Table 6 examines the effect of different text representations and tabular feature integration on recall severity prediction. Among the classical linear models, adding tabular features improved F1-score for both logistic regression and linear SVM, increasing F1 from 0.717 to 0.739, and from 0.801 to 0.816, respectively. However, logistic regression without tabular features achieved the highest recall among the TF-IDF-based models, suggesting that sparse textual features alone may provide relatively high class sensitivity but lower precision-balanced performance. Among the LightGBM models, the GloVe representation combined with tabular features achieved the best overall F1-score (0.856), indicating that the integration of dense word embeddings with structured device/regulatory information is beneficial for recall severity classification. The comparison between GloVe-only and GloVe+tabular settings also shows that tabular fea21
Table 6: Effect of text representation and tabular features on recall severity prediction. F1-score and recall are reported as macro-averaged metrics. The best value for each metric is shown in bold.
Model Logistic Regression Logistic Regression Linear SVM Linear SVM LightGBM LightGBM LightGBM LightGBM LightGBM LightGBM
Text Tabular F1 Representation Features TF-IDF No 0.717 TF-IDF Yes 0.739 TF-IDF No 0.801 TF-IDF Yes 0.816 GloVe No 0.847 GloVe No 0.844 (reason for recall only) GloVe Yes 0.856 SentenceBERT Yes 0.820 PubMedBERT Yes 0.800 BioBERT Yes 0.789
Recall 0.846 0.841 0.814 0.822 0.775 0.780 0.794 0.764 0.740 0.730
tures improve F1-score, although recall remains lower than that of TF-IDFbased logistic regression. Interestingly, transformer-derived sentence embeddings such as SentenceBERT, PubMedBERT, and BioBERT did not outperform GloVe-based LightGBM representations in this setting. This may indicate that fixed transformer embeddings, when used as static features for boosting models, do not fully exploit the contextual modeling capacity of biomedical transformers. Overall, the results suggest that both representation choice and feature integration substantially affect recall severity prediction performance, with GloVe+tabular features providing the strongest F1-score in the LightGBM setting. 5.4. Root-Cause Severity Analysis We examine the relationship between root-cause categories and recall severity through both ground-truth class distributions and model-based risk scores. The recall class distribution for each root-cause category was calculated as a percentage, and the mean ground-truth recall risk score was compared with the mean model predicted risk score. The mean ground-truth recall risk score was obtained by converting the recall class labels to the ordinal risk levels. The model-derived risk score was calculated by Equation 1 22
Table 7: Statistical and model-based root-cause recall risk analysis. Class percentages are computed within each root-cause category. Mean ground-truth (GT) recall risk is calculated using the ordinal mapping Class III = 1, Class II = 2, and Class I = 3. Mean predicted recall risk denotes the average Expected Recall Risk Score within each root-cause category. Root-Cause Category Design Human Labeling Manufacturing Material Packaging Regulatory Software Unknown
Class I (%) 10.96 2.97 1.78 9.64 8.60 5.26 2.02 1.43 6.51
Class II (%) 86.95 91.45 85.21 87.73 85.83 93.31 97.98 90.00 86.42
Class III (%) 2.09 5.58 13.02 2.63 5.57 1.42 0.00 8.57 7.06
Mean GT Recall risk 2.089 2.070 2.038 2.030 2.020 1.995 1.974 1.929 1.888
Mean Predicted Recall Risk 2.077 2.065 2.044 2.038 2.001 2.018 1.989 1.961 1.917
using the class probabilities produced by RecallRisk-BERT. As seen in Table 7, root-cause categories differ in terms of recall class distributions. While the “Design” and “Human” categories have the highest mean ground-truth recall risk scores, the “Software” and “Unknown” categories show relatively lower mean severity scores. However, the differences across categories are limited and the recall class distribution in all categories is concentrated around the dominant class. The statistical relationship between the root-cause category and recall class was evaluated using the chi-square test of independence. The test showed that there was a statistically significant relationship between root-cause category and recall class (χ2 = 316.293, p < 0.001). However, Cramér’s V was 0.121, indicating a weak effect size. This result suggests that root-cause categories are associated with recall severity, but they are not strong discriminative factors on their own. In the model-based analysis, the mean predicted recall risk was calculated for each root-cause category. The model assigned the highest predicted risk values to the “Design” and “Human” categories, whereas the lowest predicted risk values were observed for the “Unknown” and “Software” categories, as in ground-truth. The relationship between the ground-truth severity ranking and the model-based risk ranking was examined using Spearman rank correlation. The results indicated a very strong and statistically significant positive association between the two rankings (ρ = 0.983, p = 1.936 × 10−6 ). 23
This finding suggests that the model-derived recall risk scores are highly consistent with the observed root-cause severity patterns. Overall, these results indicate that RecallRisk-BERT captures meaningful relative severity patterns across root-cause categories. Nevertheless, rootcause category alone should not be interpreted as the sole determinant of recall risk. Rather, recall severity should be assessed jointly with recall narratives, device characteristics, and regulatory context, which is consistent with the proposed text–tabular multi-task modeling framework. 6. Conclusion This study develops and evaluates a post-report recall triage framework using 54,165 FDA recall records obtained from openFDA. We combine recall narratives with structured device and regulatory information, and examine both single-task recall severity prediction and multi-task learning of recall severity and root-cause category. The findings indicate that dense textual representations combined with structured tabular features provide an effective strategy for recall severity classification. We introduce the RecallRisk-BERT framework which provides a complementary multi-task approach by jointly modeling the recall severity and root-cause category. Although boosting-based models achieved the highest single-task severity performance, RecallRisk-BERT substantially improved over the single-task PubMedBERT baseline and enabled simultaneous rootcause prediction. In addition, the model-derived Expected Recall Risk Score showed strong agreement with observed root-cause severity patterns, supporting its potential use for model-based risk analysis. Overall, the findings suggest that text–tabular learning can support scalable post-report recall triage, regulatory decision support, and root-cause risk interpretation in medical device safety surveillance. The study is limited by its reliance on FDA recall records, the exclusion of complementary postmarket data sources such as adverse event reports, and the use of a single stratified train–test split. Future work should evaluate the framework across multiple random seeds, integrate additional post-market surveillance data, improve model interpretability using explainable AI techniques, and assess generalizability across international regulatory datasets.
24
References [1] Ahsan, K., & Gunawan, I. (2014). Analysis of product recalls: Identification of recall initiators and causes of recall. Operations and Supply Chain Management: An International Journal, 7(3), 97–106. [2] Barbosa Slivinskis, V., Agi Maluli, I., & Broder, J. S. (2025). A machine learning algorithm to predict medical device recall by the Food and Drug Administration. Western Journal of Emergency Medicine, 26(1), 161–170. https://doi.org/10.5811/westjem.21238 [3] Blom, T., & Niemann, W. (2022). Managing reputational risk during supply chain disruption recovery: A triadic logistics outsourcing perspective. Journal of Transport and Supply Chain Management, 16, a623. [4] Chen, W.-P., Teng, W.-G., Kuo, C. B., Yen, Y.-J., Lian, J.-Y., Sing, M., & Chen, P.-T. (2025). Regulatory insights from 27 years of artificial intelligence/machine learning–enabled medical device recalls in the United States: Implications for future governance. JMIR Medical Informatics, 13, e67552. https://doi.org/10.2196/67552 [5] FDA. (2024). Recalls, corrections and removals (devices). U.S. Food and Drug Administration. https://www.fda. gov/medical-devices/postmarket-requirements-devices/ recalls-corrections-and-removals-devices [6] Gagliardi, A. R., Takata, J., Ducey, A., Lehoux, P., Ross, S., Trbovich, P., Easty, A., Bell, C., & Urbach, D. (2017). Medical device recalls in Canada from 2005 to 2015. International Journal of Technology Assessment in Health Care, 33(6), 708–714. [7] Hu, Y., Monticolo, D., & Ghadimi, P. (2025). A machine learning-based medical device recall initiator prediction framework: From supply chain risk management and resilience view. Expert Systems with Applications. [8] Marucheck, A., Greis, N., Mena, C., & Cai, L. (2011). Product safety and security in the global supply chain: Issues, challenges and research opportunities. Journal of Operations Management, 29(7–8), 707–720. [9] M.J., A. P., T., S. K., & R., K. (2024). A comprehensive analysis of Class I medical device recalls: Unveiling patterns, causes and global impacts. Cureus, 16(8), e67542. https://doi.org/10.7759/cureus.67542 25
[10] Mukherjee, U. K., & Sinha, K. K. (2018). Product recall decisions in medical device supply chains: A big data analytic approach to evaluating judgment bias. Production and Operations Management, 27(10), 1790– 1816. [11] Sarkissian, A. (2018). An exploratory analysis of U.S. FDA Class I medical device recalls: 2014–2018. Journal of Medical Engineering & Technology, 42(8), 595–603. [12] Taylor, N. P. (2023, January 26). FDA Class I medical device recalls hit five-year high in 2022. MedTech Dive. [13] Thirumalai, S., & Sinha, K. K. (2011). Product recalls in the medical device industry: An empirical exploration of the sources and financial consequences. Management Science, 57(2), 376–392. [14] Villarraga, M. L., Guerin, H. L., & Lam, R. C. (2007). An analysis of FDA medical device recalls. Journal of Clinical Engineering, 32(2), 79–82. [15] Zhang, D., & Shen, D. (2012). Multi-modal multi-task learning for joint prediction of multiple regression and classification variables in Alzheimer’s disease. NeuroImage, 59(2), 895–907. [16] Money, A. G., Barnett, J., Kuljis, J., Craven, M. P., Martin, J. L., & Young, T. (2011). The role of the user within the medical device design and development process: medical device manufacturers’ perspectives. BMC Medical Informatics and Decision Making, 11, 1–12. [17] Ocampo, J. U., & Kaminski, P. C. (2019). Medical device development, from technical design to integrated product development. Journal of Medical Engineering & Technology, 43(5), 287–304. [18] Park, C. W., Seo, S. W., Kang, N., Ko, B., Choi, B. W., Park, C. M., ... & Yoon, H. J. (2020). Artificial intelligence in health care: current applications and issues. Journal of Korean Medical Science, 35(42). [19] Mak, K. K., Wong, Y. H., & Pichika, M. R. (2024). Artificial intelligence in drug discovery and development. Drug Discovery and Evaluation: Safety and Pharmacokinetic Assays, 1461–1498. 26
[20] Briganti, G., & Le Moine, O. (2020). Artificial Intelligence in Medicine: Today and Tomorrow. Frontiers in Medicine, 7:27. doi:10.3389/fmed.2020.00027 [21] Badnjević, A., Avdihodžić, H., & Gurbeta Pokvić, L. (2021). Artificial intelligence in medical devices: Past, present and future. Psychiatria Danubina, 33(suppl 3), 101–106. [22] Muehlematter, U. J., Daniore, P., & Vokinger, K. N. (2021). Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015–20): a comparative analysis. The Lancet Digital Health, 3(3), e195–e203. [23] Joshi, G., Jain, A., Araveeti, S. R., Adhikari, S., Garg, H., & Bhandari, M. (2024). FDA-approved artificial intelligence and machine learning (AI/ML)-enabled medical devices: an updated landscape. Electronics, 13(3), 498. [24] Clusmann, J., Kolbinger, F. R., Muti, H. S., Carrero, Z. I., Eckardt, J. N., Laleh, N. G., ... & Kather, J. N. (2023). The future landscape of large language models in medicine. Communications Medicine, 3(1), 141. [25] Chen, W., Teng, W., Kuo, C. B., Yen, Y., Lian, J., Sing, M., Chen, P. (2025). Regulatory Insights From 27 Years of Artificial Intelligence/Machine Learning–Enabled Medical Device Recalls in the United States: Implications for Future Governance. JMIR Medical Informatics, 13(1), e67552. [26] M J, A. P., Kumar, S., Kamaraj, R. (2024). A comprehensive analysis of Class I medical device recalls: Unveiling patterns, causes and global impacts, Cureus, 16(8), e67542. [27] Everhart, A.O., Sen, S., Stern, A.D., Zhu, Y., Karaca-Mandic, P. (2023). Association between regulatory submission characteristics and recalls of medical devices receiving 510(k) clearance, Journal of the American Medical Association (JAMA), 329(2), 144–156. [28] Zhalechian, M., Saghafian, S., Robles, O. (2024). Harmonizing safety and speed: A human-algorithm approach to enhance the FDA’s medical 27
device clearance policy, arXiv preprint. https://arxiv.org/abs/2407. 11823 [29] Zhu, Y., Sen, S., Everhart, A., Karaca-Mandic, P. (2025). A deep learning approach for predicting FDA’s 510(k) medical device recalls using device citation relationships, Information Systems Research [30] Hu, Y. (2024). In-depth analysis of recall initiators of medical devices with a Machine Learning–Natural Language Processing workflow. arXiv preprint arXiv:2406.10312. [31] Sarica, S., & Luo, J., 2021. Stopwords in technical language processing. PLOS ONE, 16(8), e0254937. [32] Balakrishnan, V., & Lloyd-Yemoh, E., 2014. Stemming and lemmatization: A comparison of retrieval performances. Lecture Notes on Software Engineering, 2(3), 262. [33] Singh, K. N., Devi, S. D., Devi, H. M., & Mahanta, A. K. (2022). A novel approach for dimension reduction using word embedding: An enhanced text classification approach. International Journal of Information Management Data Insights, 2(1), 100061. [34] Shi, Y., Yang, Y., & Liu, Y. (2018). Word embedding representation with synthetic position and context information for relation extraction. In 2018 IEEE International Conference on Big Knowledge (ICBK) (pp. 106–112). IEEE. [35] Bouke, M. A., & Abdullah, A. (2023). An empirical study of pattern leakage impact during data preprocessing on machine learning-based intrusion detection models reliability. Expert Systems with Applications, 230, 120715. [36] U.S. Government Accountability Office (GAO). (2025). Medical device recalls: HHS and FDA should address limitations in oversight of recall process (GAO-26-107619). https://www.gao.gov/products/ gao-26-107619 [37] LaValley, M. P. (2008). Logistic regression. Circulation, 117(18), 2395–2399. 28
[38] Joachims, T. (2002). Support vector machines. In Learning to classify text using support vector machines (pp. 35–44). Boston, MA: Springer US. [39] Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. [40] Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD (pp. 785–794). [41] Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., ... & Liu, T. Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30. [42] Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., & Gulin, A. (2018). CatBoost: Unbiased boosting with categorical features. Advances in Neural Information Processing Systems, 31, 6639–6649. [43] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. [44] Schuster, M., & Paliwal, K. K. (1997). Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11), 2673–2681. [45] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pretraining of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT (pp. 4171–4186). [46] Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. ICLR. [47] Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer. [48] Hu, Y., Monticolo, D., & Ghadimi, P. (2026). A machine learning-based medical device recall initiator prediction framework: From supply chain risk management and resilience view. Expert Systems with Applications, 298, 129922. [49] Emakhu, J., Aguwa, C., Monplaisir, L., Arslanturk, S. (2020). Failure type prediction in software-related medical device recalls. Wayne State University. In Proceedings of IISE Annual Conference, (pp. 1-6). 29
[50] Ngiam, J., Khosla, A., Kim, M., Nam, J., Lee, H., and Ng, A. Y. (2011). Multimodal Deep Learning. In Proceedings of ICML, (pp. 689–696). [51] Y. Gu, R. Tinn, H. Cheng, M. Lucas, N. Usuyama, X. Liu, T. Naumann, J. Gao, H. Poon. (2021). Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare, 3(1), 1–23. [52] J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C.H. So, J. Kang. (2020) BioBERT: A pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4), 1234–1240. [53] T. Li, W. Zhu, W. Xia, L. Wang, W. Li, P. Zhang. (2024). Research on adverse event classification algorithm of da Vinci surgical robot based on Bert-BiLSTM model. Frontiers in Computational Neuroscience, 18, 1476164. [54] A. Luschi, P. Nesi, E. Iadanza. (2023). Evidence-based clinical engineering: Health information technology adverse events identification and classification with natural language processing. Heliyon, 9(11), e21723. [55] Deznabi, I., Iyyer, M., Fiterau, M. (2021). Predicting in-hospital mortality by combining clinical notes with time-series data. In Proceedings of ACL-IJCNLP, (pp. 4026–4031), [56] Huang, K., Altosaar, J., and Ranganath, R. (2019). ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission. arXiv preprint https://arxiv.org/abs/1904.05342 [57] Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5), 513–523. [58] Pennington, J., Socher, R., Manning, C. D. (2014). Glove: Global vectors for word representation. In Proceedings of EMNLP, (pp. 689–696). [59] Schuster, M. , Paliwal., K. K. (1997) Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11), 2673–2681.
30