thesis-Dynamic Prediction of Postoperative Hospital Outcomes

Dynamic Prediction of Postoperative Hospital Outcomes

Co-Supervised by: Nico Bigler

If you are interested in this topic or have further questions, do not hesitate to contact daniel.bigler@students.unibe.ch.

Background / Context

Postoperative organ dysfunction and physiological changes may alter a patient’s subsequent hospital course. This project investigates whether longitudinal perioperative information improves dynamic prediction of live hospital discharge and in-hospital death beyond current clinical status and conventional summary features. INSPIRE contains approximately 130,000 operations with longitudinal laboratory measurements, perioperative vital signs and hospital outcomes. Published measurements are partly aggregated and quantized, so data availability and outcome definitions require validation. Access requires PhysioNet credentialing and the dataset agreement.

Research Question(s) / Goals

The research aims to evaluate dynamic postoperative prediction by:

Comparing current-state and summary-feature baselines with a sequential deep-learning model
Testing whether longitudinal history improves prediction of live hospital discharge and death
Investigating the additional value of renal-dysfunction information
Evaluating discrimination, calibration and robustness across prediction horizons

Approach / Methods

The student will:

Use INSPIRE v1.4.2 and independently audit eligible operations, longitudinal measurements, hospital outcome times and missingness
Define a postoperative cohort and a clinically reviewed creatinine-based renal-dysfunction representation
Initially predict at a fixed postoperative landmark, with a proposed seven-day horizon, among patients still hospitalized
Use only information available by the prediction time and explicitly handle incomplete follow-up
Compare logistic regression/LightGBM and a classical landmark or competing-risk survival baseline with one LSTM/GRU-based dynamic survival model
Evaluate AUPRC, AUROC, Brier score and calibration; use appropriate competing-risk evaluation for discharge and death
Assess whether temporal modelling adds value beyond current organ state and conventional summaries
Keep all records from the same subject within one data split and prevent future-information leakage

Expected Contributions / Outcomes

A reproducible dynamic prediction pipeline for postoperative hospital outcomes
A fair comparison of conventional and deep-learning-based approaches
Evidence on whether longitudinal information improves prediction, including a valid negative result
A documented assessment of missingness, outcome definitions and computational feasibility

Required Skills / Prerequisites

Python and machine learning; PyTorch and basic sequence modelling for the full Master project
Interest in clinical time series and survival analysis
Clinical supervision for outcome and phenotype definitions
For a Bachelor thesis, prepared authorized data and a reduced comparison of fixed-landmark prediction models

Possible Extensions

Additional postoperative landmarks or prediction horizons
Prediction of renal-dysfunction progression
Addition of a second organ system after validating its measurements
External validation in a compatible perioperative dataset
Transfer to a larger institutional perioperative dataset with more detailed longitudinal measurements