Sticks and stones may break your bones: Explainable methods for fracture risk prediction.

Licentiate thesis, 2025

Explaining and attributing a prediction to the model inputs is an important but often a trivial concern using classical statistical methods. In particular in medicine, where explaining a diagnosis or prognosis in terms of risk factors is crucial to understanding the condition itself, modern methods must support this kind of attributability to gain wide-spread usage. Modern neural network based approaches are often perceived as black boxes, where the exact mode of reasoning is unknown.

This work explores interpretable methods to explain risk in terms of risk factors, using survival analysis. Specifically, this thesis investigates the application of these methods to the clinical context of predicting osteoporosis-induced fractures. Osteoporosis, a serious condition associated with deteriorating bone structure and porous bones prone to fracture, is treatable but hard to diagnose at an early stage. While medical imaging is an important part of the assessment, analysis is complex and requires expert readers, and the images themselves cannot be included in the classical methods for risk prediction.

This thesis covers four papers addressing explainable methods for detecting vertebral fractures and predicting the risk of future incident fractures. Using simple, explainable algorithms, we are able to classify vertebral fractures in noisy DXA images using clinical guidelines and also demonstrate that these results are useful for downstream risk prediction. Moreover, we develop methods to predict fracture risk without annotation, attributing the risk to individual vertebrae in DXA images and different tissues in HR-pQCT images. Extensive experimental evaluations and comparisons demonstrate that these explainable methods perform better or on par with state-of-the-art, indicating that this direction of research has potential beyond interpretability