From Personal Observations to Recommendation of Tailored Interventions based on Causal Reasoning: a case study of Falls Prevention in Elderly Patients

•A novel approach for recommending tailored interventions to prevent falls using CBN.•This is the first study comparing causality-based methods to machine learning methods.•A 44-node CBN representing the most significant influences of falls among elderly.•CBN is efficient in recommending tailored interventions resulting in optimal outcome.•Classical learning methods fail to infer a precision medicine interventional paradigm. While the challenge of estimating the efficacy of therapies using observational data has received a lot of attention, little work has been done on estimating the treatment effect from interventions. In this paper, we tackle this problem by proposing an early guidance system based on a causal Bayesian network (CBN) for recommending personalized interventions. We are interested in the elderly fall prevention context. The objective is to develop a practical tool to help doctors estimate the effects of each intervention (or compound interventions) on a given patient and then choose the one that best fits each patient’s health situation to reduce the risk of falling. On a real-world elderly information base, we undertake an empirical investigation for the proposed approach, which is based on a 44-node CBN. Then, we describe what is possible to achieve using state-of-the-art machine learning methods, namely Support Virtual Machine (SVM), Decision Tree (DT), and Bayesian Network (BN), and how well these methods can be used in recommending personalized interventions compared to the proposed approach. 1174 elderly patients from Lille University Hospital, between January 2005 and December 2018 are included. The results reveal that none of the classifiers is significantly superior to the others, even if BN delivers somewhat better results (41.6%) and DT most often slightly lower performance (31.2%). Results also show that none of these three classifiers performs comparable to the proposed system (89.7%). The interventions recommended by the system are in agreement with the expert’s judgment to a satisfactory level. The reaction of the physicians to the proposed system in its first trial version was very favorable. The study showed the efficacy and utility of the causality-based strategy in recommending tailored interventions to prevent elderly falls compared to automated learning methods that had failed to infer a solid interventional paradigm for precision medicine.