Specification sensitivity of physiological clustering in intensive care: a retrospective methodological case study

Background Physiological clusters are often evaluated by resampling stability and outcome differences. These checks answer different questions from robustness to measurement representation. We examined whether a reproducible intensive-care partition remained consistent under defensible changes in observation eligibility, renal representation and numerical estimation. Methods This retrospective methodological case study used MIMIC-IV, including 55,125 four-hour periods from 9,040 adults with coded diabetes, hypertension and early physiological stress. A three-component Gaussian mixture used arterial pressure, heart rate, recorded urine output and creatinine. Patient-disjoint analyses evaluated resampling, observation rules and absolute versus relative creatinine. A post-rejection amendment fitted 90 multistart seed models across nine specifications, selecting solutions by training likelihood within each specification. A planned quantile-spline prognostic comparator was rerun after identifying a uniform-knot implementation deviation. An eICU audit evaluated external measurement feasibility. Results Fixed-preprocessing bootstrap agreement had median adjusted Rand index (ARI) 0.944. After multiseed estimation, absolute and relative creatinine partitions in the same 9,232 test periods had ARI 0.377 (conditional patient-bootstrap 95% interval 0.351 to 0.404). Selecting higher-likelihood solutions also weakened the earlier apparent freshness robustness of absolute creatinine in the reference-available subgroup. In 1,441 twelve-hour test patients with 157 deaths, adding state to the corrected source-physiology model changed AUROC by ‑0.00068 (95% interval ‑0.00565 to 0.00438). Across 208 eICU hospitals, nine had no records meeting the urine-label screen; measurement equivalence remained unresolved. Conclusions High resampling stability did not establish specification invariance. The observed partitions were associated with risk, but did not demonstrate added discrimination in the tested physiological comparator or establish clinical response phenotypes. The contribution is a reproducible, bounded demonstration of the different evidence supplied by numerical stability, representation robustness and clinical validation.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11261786/v1
Primary Topic
Machine Learning in Healthcare
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Specification sensitivity of physiological clustering in intensive care: a retrospective methodological case study

Tianyi Yu
Research Square
Machine Learning in Healthcare
preprint

Specification sensitivity of physiological clustering in intensive care: a retrospective methodological case study

Tianyi Yu
preprint en

Abstract

Background Physiological clusters are often evaluated by resampling stability and outcome differences. These checks answer different questions from robustness to measurement representation. We examined whether a reproducible intensive-care partition remained consistent under defensible changes in observation eligibility, renal representation and numerical estimation. Methods This retrospective methodological case study used MIMIC-IV, including 55,125 four-hour periods from 9,040 adults with coded diabetes, hypertension and early physiological stress. A three-component Gaussian mixture used arterial pressure, heart rate, recorded urine output and creatinine. Patient-disjoint analyses evaluated resampling, observation rules and absolute versus relative creatinine. A post-rejection amendment fitted 90 multistart seed models across nine specifications, selecting solutions by training likelihood within each specification. A planned quantile-spline prognostic comparator was rerun after identifying a uniform-knot implementation deviation. An eICU audit evaluated external measurement feasibility. Results Fixed-preprocessing bootstrap agreement had median adjusted Rand index (ARI) 0.944. After multiseed estimation, absolute and relative creatinine partitions in the same 9,232 test periods had ARI 0.377 (conditional patient-bootstrap 95% interval 0.351 to 0.404). Selecting higher-likelihood solutions also weakened the earlier apparent freshness robustness of absolute creatinine in the reference-available subgroup. In 1,441 twelve-hour test patients with 157 deaths, adding state to the corrected source-physiology model changed AUROC by ‑0.00068 (95% interval ‑0.00565 to 0.00438). Across 208 eICU hospitals, nine had no records meeting the urine-label screen; measurement equivalence remained unresolved. Conclusions High resampling stability did not establish specification invariance. The observed partitions were associated with risk, but did not demonstrate added discrimination in the tested physiological comparator or establish clinical response phenotypes. The contribution is a reproducible, bounded demonstration of the different evidence supplied by numerical stability, representation robustness and clinical validation.

Research Square
King's College London (GB)
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