Harnessing Machine Learning to Evaluate Microscopic Characteristics of Skeletal Trauma

Discriminating perimortem versus postmortem skeletal fractures remains challenging when macroscopic features overlap, particularly because fracture morphology reflects both surface characteristics and underlying tissue integrity changes. This proof-of-concept study evaluated whether machine learning models could classify fracture surfaces by integrating scanning electron microscopy (SEM) image features with quantitative tissue-integrity variables. Human femur shafts (n = 34) from 20 donors were heated under controlled conditions to simulate postmortem intervals up to 16,000 accumulated degree hours (ADH), fractured using a three-point bending setup, and imaged by SEM. The analysis used 887 SEM images grouped into perimortem (0 ADH), early postmortem (1000-3000 ADH), and late postmortem (6,000-16000 ADH) classes. Three modeling strategies were compared: tissue-integrity-only classification; SEM image-only deep learning; and a hybrid model integrating both data streams. In a fixed-configuration repeated-trial analysis, the hybrid model achieved higher mean accuracy (80%) than the tissue-integrity-only (60%) or SEM image-only (54%) models. Broader comparisons across CNN backbones and fusion classifiers showed that performance depended strongly on the model architecture and fusion strategy, with some hybrid configurations outperforming single-modality approaches. Water loss was the dominant tissue-integrity predictor, accounting for 86.6% of the variable-importance signal. Grad-CAM analysis indicated that model attention sometimes overlapped with bone-surface morphology but also included background or preparation-related regions, underscoring the need for standardized masking and additional validation. Integrating objective biomechanical measures with SEM-derived visual features provides a more accurate and scalable ML approach. However, larger donor-level studies, group-aware cross-validation, and external testing across skeletal elements, imaging conditions, laboratories, and taphonomic contexts are required before case-level application.