Deep probabilistic logic programming for diagnostic reasoning from incomplete information: A case study in stroke detection

This paper addresses the challenge of building medical diagnostic systems when raw patient data is unavailable due to privacy regulations. The authors propose a workflow that leverages summary statistics from medical literature to construct a diagnostic model. They utilize DeepProbLog, a neuro-symbolic framework that combines connectionist components (such as neural networks for detecting facial asymmetry) with probabilistic logic programming (ProbLog 2) to handle reasoning under uncertainty.

The study follows a two-step process. First, it constructs a generative probabilistic model based on causal relationships between stroke and its symptoms. Because literature data is often incomplete, the authors apply maximum entropy techniques to infer missing probabilistic dependencies. Second, they transform this generative model into a discriminative diagnostic model suitable for DeepProbLog, where the presence of a stroke is inferred from observed symptoms. To manage the complexity of the resulting model, the authors explore using ProbFOIL 2, a probabilistic inductive logic programming system, to compress the discriminative rules into a more compact and interpretable format.

By bridging the gap between deep learning-based symptom detection and rigorous probabilistic reasoning, this approach allows for the development of clinical decision support systems that are both transparent and privacy-preserving. It demonstrates how existing medical knowledge can be systematically encoded into a machine-learning architecture, reducing the need for large, sensitive, and potentially inaccessible raw datasets during the initial development phase of diagnostic tools.