The Framing of machine learning risk prediction models illustrated by evaluation of sepsis in general wards
Problem
How problem framing choices (prediction windows, observation windows, event triggers) affect the performance and clinical usability of machine learning risk prediction models, illustrated with sepsis in general wards.
Why it matters
Poorly framed sepsis alert models risk clinician fatigue from false positives and potential patient harm.
What the AI did
Applied four different framing structures to a sepsis risk prediction model on the same generic dataset and evaluated how framing affected model performance and learning.
Results
A model with strong discrimination and calibration was not necessarily clinically usable; no numeric results in abstract.
Impact
Demonstrates evaluation metrics must be assessed within framing context; no clinical deployment stated.
Record
| Venue | npj Digital Medicine |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Simon Meyer Lauritsen, Bo Thiesson, Marianne Johansson Jørgensen, Anders Hammerich Riis, Ulrick Skipper Espelund, Jesper Bo Weile, Jeppe Lange |
| Citations | 61 (OpenAlex, at import time) |
| References | 43 |
| Model or method | not stated in source |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41746-021-00529-x |
| Open access | gold |
| Keywords | machine learning, sepsis, risk prediction, problem framing, clinical usability |
Importance rating
real but narrower advance
Abstract states framing strongly affects clinician fatigue from false positives and that apparently good models may not be clinically usable; methodological advance, narrower scale.