Unsupervised machine learning for the detection and interpretation of key features in drip patterns

Abstract

Bloodstain pattern analysis (BPA) is increasingly shifting towards more objective methodologies for pattern classification. This transition can involve image-processing techniques that extract observable bloodstain features as data for pattern classification. This paper explores how unsupervised machine learning (ML)-based frameworks can be designed to identify observable features in bloodstain patterns, starting with a basic drip pattern. A total of 398 laboratory-generated drip patterns were analyzed, spanning dripping heights between 25 and 100 cm and droplet counts ranging from 1 to 10. The extracted observable features incorporated key bloodstain properties commonly used in forensic analysis, such as size and shape, hence aligning with previously reported qualitative properties and existing bloodstain taxonomies. To assess feature importance, SHAP (SHapley Additive exPlanations) analysis was applied, ranking features by their contributive power to the model’s predictions. The results revealed that the circularity, the mean intensity, and the area of the parent stain were the three most significant features for distinguishing drip patterns with contribution power of 60 %, 28 %, and 28 %, respectively, when excluding the dripping height and the number of droplets from the model. This unsupervised ML-driven approach demonstrates strong potential for establishing feature criteria for image-processing based bloodstain pattern classification methods.