Beyond AUROC: linking predictions to clinical workflows in deployed decision support systems

Background Clinical decision support systems (CDSS) require an assessment of predictive performance and the pathway supplying inputs, routing alerts, and recording workflow responses. Aggregate monitoring may not identify where predictions fail, while recorded digital interactions do not establish clinical appropriateness or patient benefit. Objective To determine whether linking records for each prediction across input data, alert decisions, delivery, clinician interaction, and outcomes identifies reliability signals missed by model performance alone. Methods We retrospectively evaluated deidentified EHR records and operational logs for four CDSS workflows: sepsis warning, inpatient deterioration, 30-day readmission, and outpatient laboratory-risk follow-up. The systems used three tree-based classifiers and one L2-regularized logistic regression. Data included 4,200 patients, 6,008 encounters, 9,402 episodes, 17,768 predictions, and 2,950 routing records. Measures before delivery covered silent validation, early deployment, and mature deployment; delivery and interaction outcomes were restricted to clinician-facing phases. Predictions were linked through input availability and timing, risk estimation, alert decisions, routing or delivery, interactions, and structured outcomes. Record review included 427 episodes. Results AUROC changes showed no consistent direction across systems (0.692–0.789); System C declined from 0.774 during silent validation to 0.692 during mature deployment. Calibration slopes ranged from 0.511 to 0.823, calibration-in-the-large from −0.633 to 0.022, and observed-to-expected ratios from 0.674 to 1.013. Early deployment had lower feature availability and longer ingestion latency. Across combinations of workflow and phase, 67.2%–87.8% of predictions had at least one unavailable feature, 21.5%–40.6% had an ingestion event after prediction, and routing or delivery failure was 2.1%–6.3%. During clinical deployment, 95.6% of production alerts were delivered; opening, acknowledgment, override, and recorded action rates were 63.8%–75.2%, 48.0%–59.0%, 10.1%–18.5%, and 17.3%–29.0%. None of the 160 subgroup contrasts remained supported after multiplicity adjustment. Conclusion Linking records for each prediction located reliability signals across input data, routing or delivery, and recorded responses that were not apparent from discrimination or performance at the operating threshold alone. Evaluation of the complete workflow complements rather than replaces discrimination, calibration, reference outcome assessment, and clinical utility evaluation. This retrospective single-institution study did not estimate causal effects or patient benefit; its estimates are local, and the monitoring structure is not a validated universal framework.

Authors

Institutions

Publication Details

Journal
Frontiers in Digital Health
Published
2026-11-18
DOI
https://doi.org/10.3389/fdgth.2026.1960268
Primary Topic
Machine Learning in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Beyond AUROC: linking predictions to clinical workflows in deployed decision support systems

Georgy Kopanitsa, Evgeny Karpulevich, Oleg Metsker, N. K. Shelukhanov et al.
1 links
Frontiers in Digital Health
Machine Learning in Healthcare
article

Beyond AUROC: linking predictions to clinical workflows in deployed decision support systems

Georgy Kopanitsa, Evgeny Karpulevich, Oleg Metsker, N. K. Shelukhanov, Yury Vitalievich Markin, Dmitry Kurapeev, Yakovlev Alexey
article en

Abstract

Background Clinical decision support systems (CDSS) require an assessment of predictive performance and the pathway supplying inputs, routing alerts, and recording workflow responses. Aggregate monitoring may not identify where predictions fail, while recorded digital interactions do not establish clinical appropriateness or patient benefit. Objective To determine whether linking records for each prediction across input data, alert decisions, delivery, clinician interaction, and outcomes identifies reliability signals missed by model performance alone. Methods We retrospectively evaluated deidentified EHR records and operational logs for four CDSS workflows: sepsis warning, inpatient deterioration, 30-day readmission, and outpatient laboratory-risk follow-up. The systems used three tree-based classifiers and one L2-regularized logistic regression. Data included 4,200 patients, 6,008 encounters, 9,402 episodes, 17,768 predictions, and 2,950 routing records. Measures before delivery covered silent validation, early deployment, and mature deployment; delivery and interaction outcomes were restricted to clinician-facing phases. Predictions were linked through input availability and timing, risk estimation, alert decisions, routing or delivery, interactions, and structured outcomes. Record review included 427 episodes. Results AUROC changes showed no consistent direction across systems (0.692–0.789); System C declined from 0.774 during silent validation to 0.692 during mature deployment. Calibration slopes ranged from 0.511 to 0.823, calibration-in-the-large from −0.633 to 0.022, and observed-to-expected ratios from 0.674 to 1.013. Early deployment had lower feature availability and longer ingestion latency. Across combinations of workflow and phase, 67.2%–87.8% of predictions had at least one unavailable feature, 21.5%–40.6% had an ingestion event after prediction, and routing or delivery failure was 2.1%–6.3%. During clinical deployment, 95.6% of production alerts were delivered; opening, acknowledgment, override, and recorded action rates were 63.8%–75.2%, 48.0%–59.0%, 10.1%–18.5%, and 17.3%–29.0%. None of the 160 subgroup contrasts remained supported after multiplicity adjustment. Conclusion Linking records for each prediction located reliability signals across input data, routing or delivery, and recorded responses that were not apparent from discrimination or performance at the operating threshold alone. Evaluation of the complete workflow complements rather than replaces discrimination, calibration, reference outcome assessment, and clinical utility evaluation. This retrospective single-institution study did not estimate causal effects or patient benefit; its estimates are local, and the monitoring structure is not a validated universal framework.

Frontiers in Digital HealthVol. 8
ITMO University (RU), Federal Almazov North-West Medical Research Centre (RU), Institute for System Programming (RU), Ministry of Health of the Russian Federation (RU)
Good health and well-being
Machine Learning in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon