Making process trace classification more explainable: approaches and experiences in the medical field

Medical process traces store the sequence of activities performed on a patient, during the implementation of a diagnostic or treatment procedure. Medical process trace classification can be used to verify whether single traces meet some expected criteria, or to make predictions about the future of a running trace, thus supporting resource planning and quality assessment. State-of-the-art process trace classification resorts to deep learning, which is proving powerful in several application domains; however, deep learning classification results are typically not explainable, an issue which is particularly relevant in medicine. In our recent work we are tackling this problem, by proposing different approaches. On the one hand, we have defined trace saliency maps, a novel tool able to visually highlight what trace activities are particularly significant for the classification task, resorting to artificial perturbations of the trace at hand that are classified in the same class as the original one. Trace saliency maps can also be paired to the corresponding countermaps, built from artificial trace perturbations classified in a different class, able to refine the map information. On the other hand, we are adopting a string alignment strategy to verify what activities are conserved, in like-nearest-neighbours (i.e., referring to the nearest traces classified in the same class as the trace under examination). Specific activities, identified in similar trace positions, can in fact justify the classification outcome, which was based on non-explainable latent features, extracted by the deep learning technique. In the chapter, we will describe the two approaches, and provide some experimental results in the field of stroke patient management.