- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Episodic Context Memory (ECM) for patient adaptation. A frozen population ECG encoder is combined with a trained multi-slot memory that retrieves prior context and corrects predictions. At deployment, unseen patients are processed chronologically without target labels or gradient updates.
Critical strengths
S1: The evaluation is comprehensive, covering external cohorts and additional physiological tasks. Repeated runs and comparisons with strong baselines provide a thorough assessment of performance and memory design.
S2: The experimental protocol clearly separates population pretraining, calibration, and target-patient inference. Its chronological design and non-overlapping windows also prevent future-information leakage.
Critical weaknesses
W1: ``Label-free personalization'' applies only at target inference. Training requires over one million labeled windows from 101 calibration patients, but no calibration-size analysis is provided.
W2: The experiments do not clearly isolate persistent multi-slot memory as the source of improvement. EMA, causal attention, and recent-context variants are close, and some transition-model ablations outperform the full router.
W3: Results vary across datasets: an appendix comparison gives the LSTM adapter the best F1/AUROC on MIT-BIH AF. This weakens the claim that ECM is consistently preferable.
Questions for authors
Q1: How does performance scale with the number of labeled calibration patients and episodes?
Q2: Which paired results show a reliable benefit from persistent multi-slot memory over recent-context conditioning?
8 comments
Response to question Q1 - Part 1/2
Q1: We thank the reviewer for suggesting this important experiment. We evaluated ECM under multiple calibration budgets in terms of available patients (Table 1) and episodes per patient (Table 2). ECM-S was tested against the best 3 non-ECM comparators: LSTM, fine-tuning, and EMA.
Table 1 shows the per-patient performance on the external cohorts when increasing the number of calibration patients in {10, 30, 50, 70}, keeping the full recording durations for each patient. Patients were subsampled from the original calibration seeds. With only 10 calibration patients, ECM-S raises external F1 from 63.6 to 75.7, reduces burden error from 24.0 to 10.3, and reduces false-AF rate from 22.8 to 4.7. ECM-S has the highest mean F1 and AUROC among the evaluated adaptation methods at every cohort size. Its advantage over fine-tuning is particularly large with limited calibration data. With 70 patients, ECM-S reaches 77.8 F1, close to the original full-cohort result of 78.5, and leads all four reported metrics. The results show that ECM-S retains substantial external benefits with substantially fewer labeled calibration patients.
Table 1. External patient-macro performance (%) across calibration cohort sizes using full recording durations. Scores are mean ± sample SD across five seeds.
| Calibration training | Method | AF F1 ↑ | AUROC ↑ | AF BE ↓ | nAF FA ↓ |
|---|---|---|---|---|---|
| None | Population | 63.6 ± 0.0 | 91.0 ± 0.0 | 24.0 ± 0.0 | 22.8 ± 0.0 |
| 10 patients; 96,376.6 ± 11,055.9 10-s ECGs | |||||
| ECM-S | 75.7 ± 2.1 | 91.0 ± 1.6 | 10.3 ± 2.9 | 4.7 ± 3.3 | |
| LSTM | 74.1 ± 4.3 | 89.6 ± 3.1 | 11.7 ± 6.0 | 7.0 ± 7.6 | |
| EMA | 73.4 ± 6.4 | 90.0 ± 2.5 | 11.3 ± 7.5 | 7.3 ± 6.0 | |
| Fine-tuning | 64.0 ± 3.8 | 84.4 ± 6.7 | 18.3 ± 3.5 | 9.8 ± 4.8 | |
| 30 patients; 291,348.8 ± 28,933.0 10-s ECGs | |||||
| ECM-S | 76.5 ± 5.2 | 91.7 ± 2.9 | 10.7 ± 4.5 | 6.5 ± 3.5 | |
| LSTM | 75.4 ± 4.7 | 91.1 ± 3.7 | 9.0 ± 4.5 | 5.0 ± 4.1 | |
| EMA | 74.9 ± 1.3 | 90.5 ± 1.6 | 8.4 ± 0.8 | 3.6 ± 2.0 | |
| Fine-tuning | 67.4 ± 2.8 | 86.3 ± 2.7 | 15.6 ± 2.0 | 4.6 ± 2.5 | |
| 50 patients; 511,685.4 ± 71,048.5 10-s ECGs | |||||
| ECM-S | 76.4 ± 1.5 | 91.8 ± 0.8 | 8.9 ± 0.7 | 2.1 ± 1.8 | |
| LSTM | 75.8 ± 3.3 | 91.4 ± 2.5 | 8.0 ± 2.5 | 3.3 ± 2.8 | |
| EMA | 75.3 ± 7.0 | 90.6 ± 3.0 | 11.1 ± 7.3 | 5.2 ± 6.1 | |
| Fine-tuning | 70.5 ± 2.7 | 87.7 ± 3.4 | 13.3 ± 2.3 | 2.8 ± 1.0 | |
| 70 patients; 735,310.8 ± 44,365.1 10-s ECGs | |||||
| ECM-S | 77.8 ± 1.4 | 92.0 ± 1.3 | 7.8 ± 0.9 | 1.4 ± 1.2 | |
| LSTM | 75.3 ± 4.6 | 87.7 ± 5.1 | 10.6 ± 3.6 | 3.4 ± 3.3 | |
| EMA | 75.6 ± 2.2 | 90.5 ± 2.0 | 9.5 ± 2.0 | 3.6 ± 1.9 | |
| Fine-tuning | 71.3 ± 4.4 | 85.5 ± 6.3 | 12.6 ± 1.8 | 3.0 ± 1.9 |
Response to question Q1 - Part 2/2
Table 2 shows the per-patient performance on the external cohorts when calibration recordings are capped at 6, 12, or 16 hours per patient.
With only 10 calibration-training patients and up to 6 hours each, ECM-S achieves 74.6 F1 versus the population model’s 63.6, and has the highest mean F1 among the evaluated methods at all three duration caps. Additional recording duration does not produce monotonic improvements in this small-cohort setting.
With the full calibration-training cohort, ECM-S’s AUROC increases from 91.3 to 91.8 to 92.4, leading the comparators at every cap. At 16 hours, it also achieves the highest mean F1 and lowest burden error, while LSTM has the lowest false-AF rate.
Therefore, substantial adaptation benefits are observed with limited calibration-training data, while benefits of additional recording duration depend on cohort size and the evaluation metric.
Table 2. External patient-macro performance (%) with 10 or 101 calibration patients and different recording-duration caps. Scores are mean ± sample SD across five seeds. Bold indicates the best mean within each condition, determined before rounding.
| Patients | Duration cap per patient | Method | AF F1 ↑ | AUROC ↑ | AF BE ↓ | nAF FA ↓ |
|---|---|---|---|---|---|---|
| 10 | 6 h (20,744.2 1,195.3 10-s ECGs) | ECM-S | 74.6 ± 1.4 | 90.1 ± 1.1 | 10.9 ± 1.8 | 7.2 ± 3.3 |
| LSTM | 74.4 ± 3.2 | 90.0 ± 1.0 | 11.7 ± 2.9 | 6.1 ± 5.4 | ||
| EMA | 72.2 ± 4.5 | 89.8 ± 1.6 | 13.4 ± 5.1 | 7.1 ± 5.1 | ||
| Fine-tuning | 59.2 ± 5.8 | 75.8 ± 7.9 | 24.1 ± 7.8 | 8.1 ± 4.8 | ||
| 10 | 12 h (40,600.0 2,779.7 10-s ECGs) | ECM-S | 75.5 ± 3.0 | 90.5 ± 1.8 | 10.6 ± 2.4 | 5.7 ± 3.7 |
| LSTM | 74.2 ± 3.8 | 90.9 ± 1.2 | 11.4 ± 4.3 | 5.9 ± 5.5 | ||
| EMA | 74.0 ± 5.1 | 90.5 ± 1.8 | 12.1 ± 6.7 | 7.6 ± 6.9 | ||
| Fine-tuning | 61.0 ± 3.3 | 81.0 ± 4.6 | 21.8 ± 4.3 | 7.7 ± 7.5 | ||
| 10 | 16 h (53,272.0$\pm$ 3,565.8 10-s ECGs) | ECM-S | 75.7 ± 2.8 | 90.7 ± 0.8 | 12.8 ± 1.4 | 6.5 ± 4.8 |
| LSTM | 74.4 ± 4.1 | 90.3 ± 1.2 | 12.5 ± 5.4 | 7.3 ± 6.4 | ||
| EMA | 73.4 ± 5.9 | 90.7 ± 2.5 | 12.2 ± 6.6 | 7.7 ± 6.3 | ||
| Fine-tuning | 57.8 ± 7.1 | 77.4 ± 10.6 | 25.4 ± 7.4 | 16.0 ± 9.1 | ||
| 101 | 6 h (163,036.8 2,209.6 10-s ECGs) | ECM-S | 76.8 ± 1.4 | 91.3 ± 0.3 | 9.2 ± 1.1 | 2.8 ± 2.7 |
| LSTM | 77.2 ± 3.2 | 89.6 ± 4.5 | 8.8 ± 1.4 | 0.8 ± 0.6 | ||
| EMA | 76.9 ± 2.8 | 90.7 ± 2.6 | 8.6 ± 0.8 | 1.3 ± 0.9 | ||
| Fine-tuning | 64.3 ± 3.4 | 78.5 ± 4.1 | 16.9 ± 2.3 | 2.8 ± 2.0 | ||
| 101 | 12 h (309,538.0 6,127.2 10-s ECGs) | ECM-S | 76.0 ± 3.1 | 91.8 ± 0.6 | 9.8 ± 2.8 | 3.2 ± 5.7 |
| LSTM | 77.5 ± 1.9 | 90.8 ± 3.3 | 8.7 ± 0.7 | 2.3 ± 2.7 | ||
| EMA | 77.2 ± 2.4 | 91.6 ± 0.9 | 8.6 ± 0.8 | 2.0 ± 1.2 | ||
| Fine-tuning | 66.8 ± 2.9 | 84.4 ± 2.8 | 15.8 ± 2.0 | 2.9 ± 1.2 | ||
| 101 | 16 h (403,418.6 7,983.4 10-s ECGs) | ECM-S | 77.7 ± 1.6 | 92.4 ± 1.0 | 7.9 ± 1.3 | 1.6 ± 2.0 |
| LSTM | 77.6 ± 4.3 | 89.8 ± 4.2 | 8.5 ± 2.0 | 1.3 ± 1.3 | ||
| EMA | 77.3 ± 2.0 | 91.5 ± 0.6 | 8.9 ± 0.4 | 3.2 ± 2.2 | ||
| Fine-tuning | 68.5 ± 3.1 | 84.3 ± 5.0 | 15.1 ± 1.4 | 4.4 ± 2.2 |
We will include the analysis in the Ablation section of the appendix.
Response to weaknesses W1-W3
We would like to thank you for the detailed assessment and constructive review. Please let us address your concerns.
W1: We thank the reviewer for pointing this out. We agree that understanding calibration cost is important to further qualify our method. As acknowledged in the Conclusion, supervised calibration enables task-aligned learning of the adaptation parameters, but simoultaneusly represents a limitation of the approach, as it introduces a further, expensive step into the pipeline. We will revise the manuscript to qualify the terminology as label-free adaptation at deployment. The requested calibration scale experiment is provided in response to Q1.
W2: We agree that the contribution of persistent memory should be distinguished from the overall benefit of patient-context adaptation. However, the simpler alternatives do not fully recover ECM’s performance: paired analyses support higher external AUROC for ECM-S and t-ECM-S than EMA and causal attention, together with prolonged monitoring improvements over EMA. Adding slow memory to recent-context conditioning improves F1 and AUROC, and removing slow memory from t-ECM-S degrades both metrics. We provide these paired comparisons under the response to Q2. Multiple slots provide separately addressable contextual representations rather than a single summary. We also explicitly report that removing the router improves mean external F1 and burden error, while the full model has slightly higher AUROC. The router was retained based on calibration-validation selection, without reselecting the architecture based on external outcomes. While this limits the evidence for routing, we believe it does not negate the improvements over simpler context models.
W3: We agree that method rankings vary across external datasets. Additional analyses help characterize this heterogeneity. 97.4% of SHDB-AF patients fall in the high embedding-distance group, compared with 34.8% of MIT-BIH AF patients. Within this group, ECM-S exceeds LSTM by 2.4% in macro-F1 (95% CI: [1.07, 3.73]) and 3.2% in AUROC ([1.64, 4.94]), whereas LSTM performs better in the medium-distance group. Separately, ECM-S’s mean F1 gain over Population is larger in the high than low waveform-distance group (19.1 vs 6.5%). These findings are therefore consistent with greater benefits under substantial measured source–deployment differences.