Machine learning in health care with a focus on breast cancer staging and intensive care prognostication

There is a never-ending need for good predictive models in medicine. Recent developments in biomarkers and imaging paired with good patient descriptions allow advanced modelling techniques to capture non-linear dependencies and complicated interactions. Advances in machine learning methods and computing power set the stage for better multivariable prediction models in breast cancer treatment and intensive care.

Preliminary results and planned studies
Study 1 A neural network model for nodal staging was externally validated in a nationwide registry cohort (n=18,633). Imputing missing values for lymphovascular invasion by machine learning predictions did not increase the model’s performance. The proportion of potentially spared
axillary interventions ranged from 16% to 24%.

Study 2 A systematic review was conducted, including published machine learning models for predicting healthy lymph nodes using imaging. Preliminary data suggests that the wide selection of imaging systems and methodologies provide sample sizes that are too small for a meta-analysis but could make a basis for a narrative synthesis.

Study 3 The successful XGBoost tree-based machine learning method will predict 90-day mortality based on admission characteristics (including novel biomarkers) of 498 critically ill COVID-19 patients. Variable importance will be assessed with SHAP, and a new bootstrap-based method for statistical analysis of SHAP values will be developed.

Study 4 The methodology of Study 3 is applied to predict the 30-day mortality of 2008 critically ill sepsis patients based on admission characteristics and novel biomarkers.