Advancing risk models: integrating blood and radiology
Our previously developed RISKINDEX predicted mortality risk from routine blood tests, age and sex. Internal development demonstrated very high accuracy, but a randomised implementation trial in real patients in the emergency department showed that doctors only acted in only 1 out of 644 cases. Adding information that clinicians do use – for example the radiology report text that accompanies an ED chest X-ray, later the image itself – will create a multimodal “RISKINDEXXR” that is both more predictive and more actionable. In this project you will build and benchmark lab-only, text-augmented and image-augmented models to predict 48-hour ICU/NIV need or 7-day mortality – endpoints that match real ED decisions better than 31-day survival.
Research question: to be determined in collaboration with supervisor
Methods for data collection:
- Data engineering: link lab outcomes and de-identified radiology reports.
- Model development: reproduce baseline; train and benchmark text-augmented models.
- Model explainability: generate SHAP / attention plots to explain new signals to clinicians.
- (Optional) Develop pilot image-fusion experiment to build image-augmented models.
Required skills: Experience with lab and radiology data; Python; Machine learning; Medical statistics.
Potential supervisor: Dr. W. van Doorn, Clinical chemist and Data & AI tech lead at Maastricht UMC+
(william.van.doorn@mumc.nl) & Dr. M. Nobel, Neuro- en Hoofd-Hals radioloog & Medical Information Officer at Maastricht UMC+ (martijn.nobel@mumc.nl)
Time period: to be agreed
Location: Maastricht/online (in consultation)