Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
Yasantha Niroshana * 1 Weijith Wimalasiri * 1 Chathuranga Hettiarachchi 1
1. Introduction
arXiv:2606.23570v1 [cs.LG] 22 Jun 2026
Abstract
Contrastive representation learning is now the default for unlabeled and weakly-labeled time-series and physiological signals (Chen et al., 2020; Le-Khac et al., 2020; Wang & Isola, 2020; Khosla et al., 2021). However, its standard formulations assume that every same-class example is equally suitable as a positive. This assumption breaks down on data where each example carries both a class label y and a subject identifier p. When class differences are partly aligned with subject differences, instance-level objectives blend class and identity together. While class-level objectives such as supervised contrastive learning fold each subject’s distinct physiology into a single shared class cluster. The resulting embeddings look well-separated overall yet are inconsistent within each subject. Therefore, a linear probe trained on one cohort fails to transfer to another. Our position is that what generalizes across subjects is not how cleanly classes are separated overall, but how consistently each subject is represented.
Contrastive representation learning struggles on physiological signals when each subject contributes a distinct baseline pattern. If class differences overlap with subject differences, classlevel objectives such as supervised contrastive learning tend to merge per-subject structure into a single per-class cluster, removing the individual variation that a model needs to generalize to unseen patients. We study this problem in the setting of Paroxysmal Atrial Fibrillation (PAF) detection from RR-interval (RRI) sequences and propose a patient-aware contrastive objective that forms positive pairs only from same-patient, sameclass segments, preserving each patient’s own sinus rhythm (SR) baseline while still pushing the two classes apart. Examining the learned embeddings directly, our objective achieves the most consistent per-patient SR structure (cohesion 0.850 vs. 0.800 for supervised contrastive loss (SupCon) and 0.772 for binary cross-entropy (BCE)). We also identify that BCE produces the cleanest global class separation yet the most disordered per-patient structure. This is precisely why a linear probe trained on its features breaks down on unseen patients. On the IRIDIA-AF dataset, the resulting representation reaches a patientindependent Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.989 ± 0.003 with 2.6× lower seed variance than supervised contrastive baselines. These results highlight that per-subject geometric consistency, rather than global class separability, is key to robust crosspatient generalization.
We study how to build this property directly into the contrastive objective. We instantiate this question on RRI sequences for Atrial Fibrillation (AF) detection: every patient has a distinct SR baseline, AF dynamics vary across individuals, and generalizing to unseen patients is the main bottleneck. Andersen et al. reported sensitivity dropping from 98.98% to 86.04% on held-out patients (Andersen et al., 2019; De With et al., 2020; Joglar et al., 2024). Selfsupervised and contrastive methods have shown promise for Electrocardiogram (ECG) signals (Grill et al., 2020; Liu et al., 2023; Hu et al., 2025; Sun et al., 2025; Chen et al., 2025), and patient-level objectives have been studied for unsupervised ECG pre-training (Kiyasseh et al., 2021; Diamant et al., 2022). What is missing is a positive-pair construction that is simultaneously class-aware and subjectaware, so it directly targets the per-class, per-subject structure that governs cross-patient transfer. Contributions.
1
Department of Computer Science, University of Moratuwa, Sri Lanka. Correspondence to: Yasantha Niroshana <[email protected]>.
• A patient-aware contrastive objective that, for each anchor, treats only same-patient, same-class segments as positives, preserving each subject’s individual structure while still pulling apart classes.
Preprint. June 23, 2026. 0 Code - github.com/EML-Labs/pacl-rri-af
1
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
• An embedding geometry analysis that explains the mechanism: our objective attains the highest perpatient SR cohesion among the compared losses. We also identify that BCE gives the cleanest global class separation yet the most disordered per-patient structure. This is consistent with its weaker transfer to new patients.
AFib Normal Normal
AFib
A
B
C
Figure 1. Standard supervised contrastive learning (left) treats all same-class segments as positives regardless of subject. The proposed patient-aware formulation (right) restricts positives to same-patient, same-class segments.
• Downstream validation on AF detection via frozenencoder linear probing on IRIDIA-AF, reaching AUROC 0.989±0.003 with 2.6× lower seed variance than supervised contrastive baselines.
2.2. Patient-Aware Mini-Batch Sampling Related work. Contrastive representation learning has progressed from instance discrimination objectives (Chen et al., 2020; Grill et al., 2020) to label-aware extensions (Khosla et al., 2021), with Wang & Isola (2020) describing the geometry of the learned representations through alignment and uniformity on the hypersphere. In the cardiac domain, several patient-level contrastive variants exist: CLOCS (Kiyasseh et al., 2021) contrasts ECG segments across space, time, and patients but is unsupervised and treats any same-patient pair as a positive; PCLR (Diamant et al., 2022) uses same-patient recordings as positives in a SimCLR-style instance task; PMQ (Sun et al., 2025) adds a patient memory queue to enrich intra-patient comparisons during pre-training. All three are unsupervised pre-training methods that use patient identity to define positives without using class labels. Our objective, in contrast, is a supervised one whose positive set is the intersection of class label and patient identity—it relies on class supervision to prevent supervised contrastive learning from collapsing each subject into a shared class cluster, while using subject identity to prevent unsupervised methods from blending class and identity. To our knowledge, no prior work applies this combined class-and-subject construction to PAF detection or analyses the resulting embedding geometry at the per-subject level.
Each mini-batch samples P patients, each contributing n SR and n AF windows (B = 2nP ). A patient is eligible only if both its SR and AF pools contain at least n windows, guaranteeing intra-patient, intra-class positive pairs in every batch and enforcing class balance across patients. 2.3. A Patient-Aware Contrastive Objective We formulate the objective as a generic template for subjectstructured data, where each sample carries a class label y and a subject identifierp. The positive set is defined to be intra-class and intra-subject. SR/AF (PAF detection) is the instantiation used here. The construction applies wherever samples are grouped by subject and the withinsubject variation is informative. For an anchor i with class yi and subject pi , the positive P(i) and negative N (i) sets are,
P(i) =
j ̸= i pj = pi ∧ yj = yi ,
(1)
N (i) =
j ̸= i yj ̸= yi .
(2)
Critically, same-class segments from different subjects are excluded from P(i), preventing the encoder from collapsing distinct individual SR baselines into a single shared prototype. This is the key departure from supervised contrastive learning (Khosla et al., 2021) (Figure 1). The per-anchor loss is the standard InfoNCE formulation, P j∈P(i) exp(sij ) P Li = − log P , (3) j∈P(i) exp(sij ) + k∈N (i) exp(sik )
2. Methodology 2.1. Dataset and Preprocessing We use IRIDIA-AF (Gilon et al., 2023), comprising longterm single-lead ECG recordings from 167 patients with paroxysmal AF. Episodes are retained when AF duration ≥ 1 hr and the immediately preceding SR duration ≥ 4 hr. Splits are patient-level (119/24/24 train/val/test); after quality filtering, 154/27/24 episodes remain. RRI sequences are RobustScaler normalized per-patient using the first SR hour as the fit window, with classification windows drawn from a strictly disjoint subsequent hour (Appendix A and Figure 3 therein). The normalized stream is segmented with a sliding window of W = 200 beats and stride S = 50 beats. Physiologically implausible beats (< 200 ms or > 2000 ms) are discarded.
with sij = ẑi⊤ ẑP j /τ and learnable temperature τ ; the batch loss is L = B1 i Li . 2.4. Encoder and Training We use a lightweight multi-branch CNN backbone to demonstrate that the gains come from the loss, not from model capacity. This is to remain compatible with edge wearable hardware. The encoder consists of three parallel 1D-CNN 2
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations Table 1. Embedding-space metrics on the test set (mean ± std, 5 seeds). Per-patient SR cohesion is the primary metric. It directly measures the per-subject geometric consistency the objective targets. The bottom row reports downstream linear-probe AUROC. Bold = best per row. ↑ higher is better, ↓ lower is better.
branches (k ∈ {3, 5, 7}) with stride convolutions (channels 16, 32, 64) (Szegedy et al., 2015), Group Normalization (Wu & He, 2018), and ReLU. Branch outputs are fused by a 1×1 convolution and pooled by a softmax temporal attention module (Shashikumar et al., 2018). A two-layer MLP projects toR128 , and both projection and pooled outputs are ℓ2 -normalized to the unit hypersphere (Wang & Isola, 2020). The contrastive loss acts on the projected embedding ê. The linear probe is trained on the pooled output ẑ. All three loss functions (Proposed, SupCon, BCE) are trained with the identical patient-aware sampler described in Section 2.3. The sampler is not varied across conditions. Training uses AdamW with a cosine scheduler. We use Optuna-TPE hyperparameter optimization (Akiba et al., 2019) to select the hyperparameters on the validation split (full configuration in Appendix B).
Metric
Proposed
SupCon
BCE
SR cohesion ↑ 0.850±0.044 0.800±0.029 0.772±0.049 AF cohesion ↑ 0.846±0.048 0.921±0.035 0.955±0.019 Centroid sim. ↓ +0.238±0.200 +0.155±0.123 −0.717±0.029 Centroid dist. ↑ 1.034±0.100 1.120±0.095 1.603±0.045 Global compact. ↑ 1.192±0.124 1.313±0.465 2.427±0.226 Per-pt. compact. ↑ 3.273±0.433 2.178±1.148 6.396±0.590 AUROC ↑
0.989±0.003
0.983±0.009
0.980±0.012
than collapsing them into a shared class prototype. The AF cohesion remains balanced (0.846, vs. 0.921 SupCon, 0.955 BCE), indicating that the SR gains do not come at the cost of disordering the AF cluster. The objective preserves both classes’ per subject structure, in contrast to SupCon’s lopsided trade-off.
3. Experiments Protocol. We adopt frozen encoder linear probing, a standard representation quality readout in contrastive learning (Wang & Isola, 2020; Chen et al., 2020; Khosla et al., 2021). Probe performance reflects the embedding geometry rather than classifier capacity. All experiments are repeated over five independent seeds and reported as mean ± std on the held-out patient-independent test split.
The BCE paradox. BCE produces the most clearly separated class means in the embedding space (centroid distance 1.603, cosine similarity −0.717) and the highest global compactness ratio (2.427). However, it gives the worst downstream AUROC on unseen patients (Section 3.2). The reason becomes clear once we look inside each patient. Those well separated class means coexist with disorganized per-patient SR placement (0.772 cohesion). A linear probe trained on seen patients has no consistent SR direction to extrapolate from when it meets a new one. This is consistent with the hypothesis that what governs transfer to new subjects is per-subject consistency, not how cleanly the classes are separated overall.. BCE achieves the highest per-patient compactness ratio (6.396). However, it is only because its denominator, the within-patient spread of each class shrinks toward zero. The positions of those tight clusters then drift unpredictably from one patient to the next, which is what the linear probe actually has to follow.
3.1. Embedding Geometry We probe the embedding geometry directly. We compare the proposed objective against SupCon (Khosla et al., 2021) and BCE baselines, fixing the encoder, sampler, and probe. Only the temperature (when applicable) τ and learning rate is varied to find the best performance. We measure, • Per-patient class cohesion, the mean cosine similarity of same-patient same-class embeddings. • Global class separability via centroid distance, centroid cosine similarity, and compactness ratio
SupCon over-compacts AF, under-aligns SR. SupCon achieves the tightest AF cohesion (0.921) by pulling all AF segments together regardless of subject, at the direct cost of SR cohesion (0.800 vs. 0.850). This is the predicted failure mode. Standard supervised contrastive learning trades perpatient SR structure for global AF compactness, precisely the tradeoff that hurts patients with atypical SR baselines.
• Per-patient compactness ratio. Full definitions are in Appendix C. Results are summarized in Table 1. SR cohesion is the primary metric because sinus rhythm is each patient’s individual baseline which is the reference point a cross-patient linear probe must extrapolate from when it encounters a new subject.High AF cohesion without corresponding transfer improvement confirms that SR placement consistency, not AF compactness, is the bottleneck.
3.2. Downstream PAF Detection The geometric ordering is preserved on the downstream task (Figure 2). AUROC orders the three losses identically to per-patient SR cohesion (Section 3.1). The most striking effect is on stability. Seed to seed AUROC variance is reduced by 2.6× over SupCon and 3.4× over BCE. A geometrically consistent embedding space is less sensitive
Per-patient SR cohesion confirms the design. The proposed objective achieves the highest SR cohesion (0.850 vs. 0.800 SupCon vs. 0.772 BCE), evidence that intra subject positives preserve each subject’s own SR baseline rather 3
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations Per-metric comparison (mean ± std, 5 seeds)
1.000
by subject (EEG, EMG, PPG) remains an open empirical question and a direction for future work. On the clinical side, the consistent AF recall (0.974 ± 0.008) across seeds matters. A missed AF episode may delay anticoagulation and elevate stroke risk, so seed-stable sensitivity is a deployment advantage. The gain is in the representation, not the classifier. This downstream performance is achieved with a frozen encoder and a logistic regression probe and exceeds Andersen et al.’s patient-holdout sensitivity (Andersen et al., 2019) by over 11 percentage points (Appendix E, Table 5).
0.975
Score
0.950 0.925 0.900 0.875 0.850
Proposed Loss SupCon BCE
0.825 0.800
-AUC
ROC
ion
recis
AF P
ecall
AF R
AF F
1
SR F
1
racy
Accu
Limitations. Our analysis is on a single dataset (IRIDIAAF) with a frozen-probe protocol. Prospectively, multicentre validation and end-to-end fine-tuning remain open. Three concrete next steps follow from the geometric findings.
Figure 2. Per-metric linear-probe comparison of the proposed patient-aware objective, SupCon, and BCE on the held-out patientindependent test split (5 seeds). Bars show the seed mean and error bars show ± one standard deviation.
to weight initialization, which is a desirable property for any deployment that must reproduce model behaviour. AUROC improvements of +0.006 over SupCon and +0.009 over BCE each exceed one standard deviation of the corresponding baseline. Per-class precision/recall and a comparison to prior RRI-based AF detectors are reported in Appendix E (Tables 3 and 5). The proposed representations exceed Andersen et al.’s patient-holdout sensitivity (Andersen et al., 2019) by over 11 percentage points despite using only a frozen encoder and a logistic-regression probe.
• A per-subject decomposition of alignment and uniformity to verify the mechanism formally. • Few shot adaptation that uses each patient’s preserved SR structure as a personal prior. • Applying the same class and subject positive construction to other physiological signals where between subject variability is the main barrier to generalization.
5. Conclusion 4. Discussion
We proposed a patient-aware contrastive objective for physiological signals that are organized by subject. Positives are restricted to same-patient, same-class pairs, preserving each subject’s own structure while still pushing the two classes apart. On IRIDIA-AF, the construction attains the most consistent per-patient SR cohesion among the losses we compared, uncovers a BCE paradox. Overall class separability is a misleading indicator of how well a representation transfers to unseen subjects. Validated downstream, the construction reaches AUROC 0.989 ± 0.003 for PAF detection on unseen patients with 2.6× lower seed variance than supervised contrastive baselines. The objective does not depend on the encoder and only requires subject IDs at training time. This suggests same-class, same-subject positive construction as a broadly useful primitive for representation learning on signals that are organized by subject.
Per-subject consistency, not global separability, governs transfer. The BCE paradox is the clearest demonstration of our central claim. Class separability metrics tell us how far the class means have been pushed apart, but a linear probe trained on seen patients can only generalize to new ones if the embedding space provides a consistent direction along which to extrapolate. When the per-patient SR structure is disorganized, that direction is no longer well-defined even if the class means themselves are far apart. The proposed objective targets this consistency directly by forming positives only from same-patient, same-class segments, and its +0.050 SR-cohesion advantage over SupCon translates into a 2.6× reduction in AUROC variance across seeds. Geometrically, the loss maintains a per subject balance between alignment and uniformity. Each patient’s class-conditional cluster is tightly aligned while clusters from different patients remain spread out on the hypersphere. This avoids both SupCon’s collapse into a single shared class prototype and BCE’s cross-patient inconsistency.
Acknowledgements We thank Joshua Pranjeevan Kulasingham for his support.
Generality and clinical relevance. The construction does not depend on the encoder and only requires subject identifiers and class labels in each batch. The mechanism applies wherever (a) class labels are available at training time and (b) within-subject variation is informative. Whether the same gains transfer to other physiological signals organized
References Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M. Optuna: A Next-generation Hyperparameter Optimization Framework, July 2019. URL http://arxiv. org/abs/1907.10902. arXiv:1907.10902 [cs]. 4
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
Andersen, R. S., Peimankar, A., and Puthusserypady, S. A deep learning approach for realtime detection of atrial fibrillation. Expert Systems with Applications, 115:465–473, January 2019. ISSN 0957-4174. doi: 10.1016/j.eswa.2018.08. 011. URL https://www.sciencedirect.com/ science/article/pii/S0957417418305190.
Gheshlaghi Azar, M., Piot, B., kavukcuoglu, k., Munos, R., and Valko, M. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Advances in Neural Information Processing Systems, volume 33, pp. 21271–21284. Curran Associates, Inc., 2020. Hu, J., Li, C., Cao, J., and Kou, B. A novel multimodal self-supervised framework for ECG arrhythmia classification. Computers in Biology and Medicine, 198:111137, November 2025. ISSN 00104825. doi: 10.1016/j.compbiomed.2025.111137. URL https://www.sciencedirect.com/ science/article/pii/S0010482525014908.
Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of ICML’20, pp. 1597–1607. JMLR.org, July 2020. URL https://dl. acm.org/doi/10.5555/3524938.3525087.
Joglar, J. A., Chung, M. K., Armbruster, A. L., Benjamin, E. J., Chyou, J. Y., Cronin, E. M., Deswal, A., Eckhardt, L. L., Goldberger, Z. D., Gopinathannair, R., Gorenek, B., Hess, P. L., Hlatky, M., Hogan, G., Ibeh, C., Indik, J. H., Kido, K., Kusumoto, F., Link, M. S., Linta, K. T., Marcus, G. M., McCarthy, P. M., Patel, N., Patton, K. K., Perez, M. V., Piccini, J. P., Russo, A. M., Sanders, P., Streur, M. M., Thomas, K. L., Times, S., Tisdale, J. E., Valente, A. M., Van Wagoner, D. R., and Peer Review Committee Members. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation, 149(1):e1–e156, January 2024. ISSN 1524-4539. doi: 10.1161/CIR.0000000000001193.
Chen, W., Wang, H., Zhang, L., and Zhang, M. Temporal and spatial self supervised learning methods for electrocardiograms. Sci Rep, 15(1):6029, February 2025. ISSN 2045-2322. doi: 10.1038/s41598-025-90084-2. De With, R. R., Erküner, O., Rienstra, M., Nguyen, B.O., Körver, F. W. J., Linz, D., Cate Ten, H., Spronk, H., Kroon, A. A., Maass, A. H., Blaauw, Y., Tieleman, R. G., Hemels, M. E. W., de Groot, J. R., Elvan, A., de Melis, M., Scheerder, C. O. S., Al-Jazairi, M. I. H., Schotten, U., Luermans, J. G. L. M., Crijns, H. J. G. M., and Van Gelder, I. C. Temporal patterns and short-term progression of paroxysmal atrial fibrillation: data from RACE V. Europace, 22(8):1162– 1172, August 2020. ISSN 1099-5129. doi: 10.1093/ europace/euaa123. URL https://pmc.ncbi.nlm. nih.gov/articles/PMC7400474/.
Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., and Krishnan, D. Supervised Contrastive Learning, March 2021. URL http:// arxiv.org/abs/2004.11362. arXiv:2004.11362 [cs].
Diamant, N., Reinertsen, E., Song, S., Aguirre, A. D., Stultz, C. M., and Batra, P. Patient contrastive learning: A performant, expressive, and practical approach to electrocardiogram modeling. PLoS Comput Biol, 18 (2):e1009862, February 2022. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1009862.
Kiyasseh, D., Zhu, T., and Clifton, D. A. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients. In Proceedings of the 38th International Conference on Machine Learning, pp. 5606–5615. PMLR, July 2021. URL https://proceedings. mlr.press/v139/kiyasseh21a.html.
Faust, O., Shenfield, A., Kareem, M., San, T. R., Fujita, H., and Acharya, U. R. Automated detection of atrial fibrillation using long short-term memory network with RR interval signals. Comput Biol Med, 102: 327–335, November 2018. ISSN 1879-0534. doi: 10.1016/j.compbiomed.2018.07.001.
Le-Khac, P. H., Healy, G., and Smeaton, A. F. Contrastive Representation Learning: A Framework and Review. IEEE Access, 8:193907–193934, 2020. ISSN 2169-3536. doi: 10.1109/ACCESS.2020.3031549. URL https:// ieeexplore.ieee.org/document/9226466.
Gilon, C., Grégoire, J.-M., Mathieu, M., Carlier, S., and Bersini, H. IRIDIA-AF, a large paroxysmal atrial fibrillation long-term electrocardiogram monitoring database. Sci Data, 10(1):714, October 2023. ISSN 2052-4463. doi: 10.1038/ s41597-023-02621-1. URL https://www.nature. com/articles/s41597-023-02621-1.
Liu, W., Li, Z., Zhang, H., Chang, S., Wang, H., He, J., and Huang, Q. Dense lead contrast for selfsupervised representation learning of multilead electrocardiograms. Information Sciences, 634:189–205, July 2023. ISSN 0020-0255. doi: 10.1016/j.ins.2023.03. 099. URL https://www.sciencedirect.com/ science/article/pii/S002002552300422X.
Grill, J.-B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., 5
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
Shashikumar, S. P., Shah, A. J., Clifford, G. D., and Nemati, S. Detection of Paroxysmal Atrial Fibrillation using Attention-based Bidirectional Recurrent Neural Networks. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’18, pp. 715–723, New York, NY, USA, July 2018. Association for Computing Machinery. ISBN 978-1-4503-5552-0. doi: 10.1145/3219819. 3219912. URL https://dl.acm.org/doi/10. 1145/3219819.3219912. Sun, X., Yang, Y., and Dong, X. Enhancing Contrastive Learning-based Electrocardiogram Pretrained Model with Patient Memory Queue. May 2025. doi: 10.48550/arXiv. 2506.06310. URL http://arxiv.org/abs/2506. 06310. arXiv:2506.06310 [eess]. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, June 2015. doi: 10.1109/CVPR. 2015.7298594. URL https://ieeexplore.ieee. org/document/7298594. ISSN: 1063-6919. Udawat, A. S. and Singh, P. An automated detection of atrial fibrillation from single-lead ECG using HRV features and machine learning. J Electrocardiol, 75:70–81, 2022. ISSN 1532-8430. doi: 10.1016/j.jelectrocard.2022.07. 069. Wang, T. and Isola, P. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, pp. 9929–9939. PMLR, November 2020. URL https://proceedings.mlr.press/v119/ wang20k.html. Wu, Y. and He, K. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018.
6
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
A. Preprocessing Pipeline ≥ 4 hours
≥ 1 hour
1 hour
1 hour
1 hour
Scaler
SR
AFib
Figure 3. Episode selection and preprocessing pipeline. SR normalization windows (hour 0–1) are disjoint from classification windows (hours 1–2), ensuring the RobustScaler has no access to labeled data. The ≥ 4 hr SR inclusion criterion further ensures that all SR classification windows begin at least 2 hr before AF onset, reducing the risk of including pre-episode transitional rhythms.
Per-patient normalization uses, (p)
˜ i = RRi − median(Rscale ) RR (p) IQR(Rscale )
(4)
(p)
where Rscale is the patient specific fit window (the first SR hour). Beats with RR < 200 ms or > 2000 ms are excluded as physiologically implausible. The hyperparameter optimization is described in Appendix B.
B. Encoder Architecture and Training Details 200 RR Intervals
Feature Maps Kernel Size (k)
Concatenated Feature Maps
Output Feature Vector
Mixed Features
Latent Vector
k=3
k=5 Channel Mix (kernel size = 1 )
Attention Pooling
Projection
k=7
Figure 4. Encoder architecture: multi-branch 1D-CNN backbone with temporal attention pooling and a two-layer MLP projection head.
Architecture. Three parallel 1D-CNN branches with kernel sizes k ∈ {3, 5, 7} capture beat-to-beat fluctuations, mediumrange oscillations, and broader trend dynamics, motivated by Inception style multi-scale processing (Szegedy et al., 2015). Each branch applies three strided convolution blocks (stride 2) with channel depths {16, 32, 64}, Group Normalization (Wu & He, 2018) (8 groups), and ReLU. Branch outputs are concatenated and fused via a 1×1 convolution, then pooled by a softmax temporal attention module (Shashikumar et al., 2018), αt = P
exp(w⊤ ht ) , ⊤ ′ t′ exp(w ht )
z=
X
αt ht .
(5)
t
A two-layer MLP projects z ∈ R128 to e ∈ R128 ; both are ℓ2 -normalized to the unit hypersphere (Wang & Isola, 2020). The contrastive loss operates on ê; the linear probe on ẑ. Training. PyTorch on a single NVIDIA RTX 2080. AdamW with learning rate 6.8×10−3 and weight decay 8.8×10−4 ; cosine annealing to 10−6 ; dropout 0.12 on the projection head; learnable temperature τ initialized at 0.05. Patient-aware 7
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
sampling uses P = 8 patients and n = 16 windows per class per-patient (B = 128). Up to 100 epochs with early stopping on validation AUROC (patience 10). Hyperparameter selection. 32 hyperparameters (architecture, optimizer, data pipeline including W and S) are jointly tuned by Optuna’s TPE sampler (Akiba et al., 2019) on the validation split, with the test split held out until final evaluation. For all three loss conditions (Proposed, SupCon, BCE), only the temperature τ and learning rate are varied to find the best performance per loss. The encoder architecture, sampler, and probe are held fixed across all conditions, giving no tuning advantage to the proposed loss.
C. Embedding-Geometry Metric Definitions The encoder produces L2-normalized embeddings, so every embedding zi ∈ Sd−1 satisfies ∥zi ∥ = 1. Let yi ∈ {SR, AF} denote the class label and pi ∈ P the patient identifier. Define the index sets S c = {i : yi = c},
Spc = {i : yi = c, pi = p},
the (unnormalized) class centroids 1 X zi , |S c | c
µc =
µcp =
i∈S
1 X zi , |Spc | c i∈Sp
their L2-normalized versions µ̂c = µc /∥µc ∥, µ̂cp = µcp /∥µcp ∥, and the mean class spreads d¯c =
1 X ∥zi − µc ∥2 , |S c | c
d¯cp =
i∈S
Per-patient class cohesion (primary metric).
1 X ∥zi − µcp ∥2 . |Spc | c i∈Sp
For class c ∈ {SR, AF},
Coh c =
1 X 1 X ⊤ c zi µ̂p . |P | |Spc | c p∈P
(6)
i∈Sp
Because ∥zi ∥ = ∥µ̂cp ∥ = 1, each summand is the cosine similarity between an embedding and its same-patient same-class centroid; values lie in [−1, 1] and approach 1 as the per-patient class cluster tightens. Global class separability. CentDist = ∥µSR − µAF ∥2 ,
(7)
CentSim = µ̂SR ⊤ µ̂AF ,
(8)
Cglob =
∥µSR − µAF ∥2 . d¯SR + d¯AF
(9)
We refer to Cglob as a compactness ratio. It is the ratio of between-class centroid distance to the sum of within-class mean spreads. Per-patient compactness. Cpp =
AF 1 X ∥µSR p − µp ∥2 ¯AF . |P | d¯SR p + dp
(10)
p∈P
Cpp measures the same separation-to-spread trade-off as Cglob , but evaluated within each patient and then averaged. ¯AF Note: when within-patient class spread is very small (d¯SR p + dp → 0), the ratio becomes large regardless of whether the resulting clusters are positioned consistently across patients. In such cases Cpp should be interpreted alongside the raw spreads d¯cp and the per-patient cohesion scores rather than in isolation. 8
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations Table 2. Embedding-space metrics on the test set (mean ± std, 5 seeds). Per patient SR cohesion is the primary metric. Bold = best per row. ↑ higher is better; ↓ lower is better. Metric
Proposed
SupCon
BCE
SR cohesion ↑ AF cohesion ↑ Centroid similarity ↓ Global centroid dist. ↑ Global compactness ratio ↑ Per-patient compactness ratio ↑
0.850 ± 0.044 0.846 ± 0.048 +0.238 ± 0.200 1.034 ± 0.100 1.192 ± 0.124 3.273 ± 0.433
0.800 ± 0.029 0.921 ± 0.035 +0.155 ± 0.123 1.120 ± 0.095 1.313 ± 0.465 2.178 ± 1.148
0.772 ± 0.049 0.955 ± 0.019 −0.717 ± 0.029 1.603 ± 0.045 2.427 ± 0.226 6.396 ± 0.590
AUROC ↑
0.989 ± 0.003
0.983 ± 0.009
0.980 ± 0.012
Table 3. Per-class linear-probe metrics for the proposed representations (4160 AF and 3799 SR segments from unseen patients; mean ± std across 5 seeds). The headline AUROC of 0.989 ± 0.003 is reported in the main text. Class
Prec.
Recall
F1
SR (0) AF (1)
0.970 ± 0.009 0.936 ± 0.017
0.927 ± 0.020 0.974 ± 0.008
0.948 ± 0.012 0.955 ± 0.010
0.952 ± 0.011 0.989 ± 0.003
Accuracy AUROC
D. Embedding-Geometry Metrics (Full Table) E. Downstream PAF Detection (Per-Class Metrics and Loss Comparison) F. Comparison with Prior Work
9
Patient-Aware Contrastive Learning Preserves Per-Patient Structure in RR-Interval Representations
Table 4. Linear-probe comparison across loss functions on the fixed patient-independent test split (5 seeds; bold = best mean per metric). Loss
AUROC
AF Prec.
AF Recall
AF F1
SR F1
Acc.
Proposed 0.989±0.003 0.936±0.017 0.974±0.008 0.955±0.010 0.948±0.012 0.952±0.011 SupCon 0.983±0.009 0.933±0.017 0.972±0.005 0.952±0.008 0.945±0.011 0.949±0.009 BCE 0.980±0.012 0.929±0.021 0.974±0.009 0.951±0.012 0.943±0.016 0.947±0.014
Table 5. Performance comparison with related work on RR-interval AF detection. Direct numerical comparison is limited by dataset and evaluation differences. † Cross-validation with within-patient data mixing inflates reported metrics. ‡ MIT-BIH AF contains no episodes satisfying our quality criteria (AF ≥ 1h, preceding SR ≥ 4h). Study
Method
Dataset
Validation †
(Udawat & Singh, 2022) HRV+ML MIT-BIH AF CV (Andersen et al., 2019) CNN+RNN 3 dbs 5-fold CV† (Andersen et al., 2019) CNN+RNN Unseen Pt. holdout (Faust et al., 2018) LSTM MIT-BIH‡ 10-fold CV† Ours
CNN+MLP IRIDIA-AF
Pt. holdout
10
Sens.
Spec.
Acc.
95.16% 98.98% 86.04% —
92.46% 96.95% 98.96% —
94.43% — — 98.51%
97.40±0.80% 92.72±2.01% 95.17±1.07%