A NeuroSymbolic Human-in-the-loop Approach Towards Fusing Medical Expert Knowledge with ANNs

A NeuroSymbolic Human-in-the-loop Approach Towards Fusing Medical Expert Knowledge with ANNs AIAI 2025 - SilverTech Workshop Spyros Theodoropoulos∗,∗∗ , Georgios Makridis∗∗ , Aristodemos Pnevmatikakiss∗∗∗ , Vrettos Moulos∗, Dimosthenis Kyriazis∗∗ , Panayiotis Tsanakas∗ ∗National Technical University of Athens ∗∗ University of Piraeus ∗∗∗ Innovation Sprint Srl Introduction • The ambition of NeuroSymbolic AI (NSAI) is to fuse symbolic knowledge derived from experts with neural networks that learn from vast amounts of data • In high-stakes applications such as healthcare, NSAI offers a promising route to increase transparency and trust in automated decision systems. • The management of chronic diseases can serve as a testbed for NSAI, especially with regards to generating personalized lifestyle adjustment advice. • Our main use case is the combination of expert guidelines with model knowledge derived from Diabetes Mellitus Type 2 patients’ lifestyle data, in order to mitigate hyperglycemia symptoms and improve quality of life. 1 The Diabetes Mellitus Type 2 Dataset →Synthetic dataset provided by Innovation Sprint, based on early data from patients treated at the University Clinic of Endocrinology and Metabolic Diseases of the General University Hospital of Larisa. →Seed data collected using Healthentia, ISPs remote patient monitoring platform from a single patient over the course of 8 weeks to synthesize data over the course of 14 weeks. Activity Data Description Steps Number of steps walked by patient on specific day Calories Number of calories consumed Floors Amount of stairs climbed in floors Intensity minutes Time spent performing intense exercise Sleep The hours of sleep per day Weight The latest weight measurement of the patient 2 Prediction The following characteristics of the dataset set the requirements for the NSAI approach: •Target: Predict the patient’s SMBG value using their activity characterisics. • The SMBG value depends largely on the time of the measurement e.g. BB: Before Breakfast. • Predictions should be consistent with goals set by the doctors: Goals: 1. Walk at least 500 steps more than the baseline established in weeks -2 and -1. 2. Exercise at least 150 moderately or intensively (counting x2) minutes per week. 3. Sleep at least 6 but no more than 8 hours every night. 3 SMBG Values Figure 1: Distributions of SMBG values per meal. 4 NeuroSymbolic AI (1/2) Learning for Reasoning vs. Reasoning for Learning Learning for Reasoning: Extensions of symbolic reasoning methods that utilize empirical machine learning to handle unstructured data or accelerate reasoning. Reasoning for Learning: Symbolic knowledge used in neural classifiers, through knowledge transfer or regularization. Key approaches include Logic Tensor Networks (LTN)1and the Symbolic Probabilistic Layer (SPL)2. 1Samy Badreddine et al. “Logic Tensor Networks”. In: Artificial Intelligence 303 (2022), p. 103649. DOI: https://doi.org/10.1016/j.artint.2021.103649. 2Kareem Ahmed et al. “Semantic Probabilistic Layers for Neuro-Symbolic Learning”. In: Advances in Neural Information Processing Systems. Ed. by S. Koyejo et al. Vol. 35. Curran Associates, Inc., 2022, pp. 29944–29959. 5 NeuroSymbolic AI (2/2) Semantic Probabilistic Layer (SPL): →Enforces logical constraints via Ordered Binary Decision Diagrams (OBDDs) transformed into differentiable Probabilistic Circuits (PCs). →Readjusts probability conversions, ensuring consistency with rules. →Low sample complexity but limited to simple logical propositions (no first-order logic). Logic Tensor Networks (LTN): →Grounding of first-order logic propositions to real-valued tensors for truth degrees. →Tensors guide back-propagation through loss function regularization. →High accuracy, low sample complexity in various domains. →Constraints’ satisfiability is not fully guaranteed. 6 Towards a HIL Framework Approach We propose a Human-in-the-loop architecture where: • Medical professionals can embed their expertise into a NeuroSymbolic model by formulating human-understandable rules. • The system provides explainable AI (XAI) methods to inspect the model’s reasoning and evaluate correctness against expert knowledge. • This enables an iterative refinement process to adjust rules and improve model behavior. 7 HIL Framework Figure 2: Use case diagram of the proposed NSAI component for healthcare. 8 Conclusion and Future Work • The results presented are a promising starting point for combining expert rules with data-driven learning in intelligent healthcare applications. • The benefits of the approach are expected to become more pronounced on larger and noisier datasets, provided accurate symbolic knowledge is supplied by healthcare experts. •Immediate focus: Diabetes use case • Integrate anonymized demographic and medical examination data into feature vectors • Incorporate rules derived from questionnaires and common-sense guidelines •Future direction: Explore meta-cognitive approaches where rules are refined interactively with user input to ensure alignment with clinical insights. 15