Evaluation of latent health risk prediction models: A protocol for a mixed-methods study using clinical triage as a vehicle for comparison and discussion (Preprint)

<sec> <title>BACKGROUND</title> Determining clinical urgency and resource allocation within acute patient populations is complex. Tools are being developed to capture global assessments of patient health beyond disease-specific scores, aiming to provide dynamic assessments incorporating both baseline physiological reserve and immediate illness severity </sec> <sec> <title>OBJECTIVE</title> This study evaluates two contrasting approaches to latent health measurement: FI-lab, a transparent algorithmic tool using bottom-up aggregation of laboratory abnormalities, and ETHOS-ARES, a transformer-based model using top-down learning from electronic health records. </sec> <sec> <title>METHODS</title> This two-phase mixed-methods study uses clinical triage scenarios with ≥30 clinicians sampled across hospital roles (physicians, surgeons, critical care nurses, advanced practitioners). Phase 1 compares unaided clinician judgments of severity and clinical urgency against model outputs using Spearman's rank correlation (ρ ≥ 0.70 indicates good agreement as primary outcome). A novel "clinical Turing test" assesses whether model rankings are statistically distinguishable from clinician assessments. Phase 2 allows clinicians to incorporate model outputs into identical tasks, measuring anchoring effects through within-person pre/post comparison. Semi-structured interviews explore clinical utility, trust, and perceived limitations. Case materials derive from MIMIC-IV-ED, presented as slide decks containing emergency department documentation, examination findings, laboratory results, and imaging reports. Qualitative analysis follows the Framework Method with dual independent coding. </sec> <sec> <title>RESULTS</title> Data collection is planned for October-November 2025, with analysis to follow in December 2025. </sec> <sec> <title>CONCLUSIONS</title> We anticipate findings will quantify agreement between model outputs and clinician consensus, measure any anchoring effects from model exposure, and generate insights from qualitative data regarding clinical utility, feasibility, and factors influencing clinician trust and adoption of different approaches to latent health measurement. </sec> <sec> <title>CLINICALTRIAL</title> n/a </sec>

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.