Related Experiment Video
Updated: Oct 7, 2026
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Auditing diagnosis-conditioned electronic health record exports for secondary use: A selection-aware study in child
Seda Bayat Toksöz1, Gültekin Işık1, Eda Gülbetekin2
1Department of Computer Engineering, Faculty of Engineering, Iğdır University, Şehit Bülent Yurtseven Campus, 76000 Suveren, Iğdır, Türkiye.
Background:
Diagnosis-conditioned electronic health record (EHR) exports can embed selection effects in observed associations and machine-learning results.
Objective:
To evaluate Selection-Aware Mapping and Predictive Adequacy for Electronic Health Records (SAMPA-EHR) as a rejection-capable audit for secondary use.
Methods:
Four outpatient exports for autism, attention-deficit/hyperactivity disorder (ADHD), specific learning disorder, and anxiety (1 October 2025-21 July 2026) were merged at encounter level. The audit separated export-family selection diagnostics from positive co-coding with false-discovery-rate control, evaluated non-negative matrix factorization (NMF) ranks 2-6 across 50 initializations against stability and component-size gates, and benchmarked administrative features using repeated nested cross-validation, training-fold calibration, prevalence references, and a calendar-period stress test.
Results:
The 1,853 source rows yielded 1,631 encounters and 30 clinical labels; 421 encounters (25.8 %) had at least two labels. All six export-family associations were negative, including ADHD-anxiety despite 120 joint encounters (odds ratio 0.25; phi - 0.292), a pattern consistent with ascertainment effects. Seven positive label pairs met the descriptive retention criteria. No NMF rank passed both eligibility gates (highest mean adjusted Rand index: 0.742 at rank 2). Logistic regression showed no improvement over the prevalence reference (AUROC 0.505, 95 % CI 0.475-0.535; AUPRC 0.258 versus 0.258); calendar-period AUROC was 0.500.
Conclusions:
SAMPA-EHR distinguished reportable local co-coding from latent and predictive outputs that did not meet the audit criteria. The findings support encounter-level data-governance decisions within the evaluated extraction context; they do not establish patient-level phenotypes or clinical prediction.