Hybrid AI Approach for Counterfactual Prediction over Knowledge Graphs for Personal Healthcare [Draft]

https://youtu.be/vkF3fpCZLhk Artificial intelligence (AI) has become an invaluable tool in health-care for disease prediction and diagnosis. Despite their predictive accuracy, AI models may ignore causal relationships between patient characteristics (demographic or clinical). As a result, although AI models capture associative patterns, ignoring the causal relationships of predicted characteristics limits their ability to perform counterfactual reasoning about the predicted characteristics of a patient when dependent characteristics change. This limitation of AI models can affect the understanding of predicted outcomes. We have proposed hybrid AI methods that combine symbolic reasoning over knowledge graphs (KGs), large language models (LLMs), and causal reasoning techniques to infer causal relationships between patients' properties. As a result, a causal model is learned, enabling counterfactual prediction to support clinical decisions. We apply these AI methods to predict the counterfactuals of biomarker results in non-small cell lung cancer (NSCLC) patients under hypothetical treatments with different smoking habits. We have created synthetic datasets based on clinical records of NSCLC patients to evaluate the performance of the proposed methods. The observed results suggest that our methods are competitive with baseline methods in causal relationship discovery and counterfactual prediction. CCS CONCEPTS • Applied computing → Health care information systems; • Computing methodologies → Artificial intelligence.