Determine the cause of rare same-subject retrieval enrichment

Determine why same-subject retrieval occurs slightly more often than expected under random retrieval in the training fold of PRISM’s dual-modality population memory bank.

Background

PRISM retrieves neighboring encounters from separate EHR- and ECG-based memory-bank indices. Although evaluation partitions are patient-disjoint, patients with multiple encounters can in principle retrieve one of their own encounters within the training fold. The observed same-subject retrieval rate is very small but exceeds the approximate random baseline, and the paper does not establish the reason for this enrichment.

References

This remains a very small fraction despite 76,875 rows in the combined training and validation cohort (62.7% of 122,641 encounters) belonging to patients with multiple encounters; the cause of this small enrichment is not established.

— Retrieval-Augmented Multimodal Language Model with Heterogeneous Evidence Integration for Clinical Prognosis  (2610.13510 - Shafiq et al., 9 Oct 2026) in Section Population Memory Bank, paragraph 'Same-patient retrieval'