Bayesian Optimization for Clinical Pathway Decomposition From Aggregate Data

Data protection rules often impose anonymization requirements on datasets by means of aggregations that hinder the exact simulation of individual subjects. For example, clinical pathways that disclose medical conditions of patients may typically need to be aggregated to preserve anonymity of the subjects. However, aggregation unavoidably results in biasing the simulation process, for example, by introducing spurious pathways that can skew the simulated trajectories. In this paper, we study this problem and develop approximate decomposition methods that mitigate its impact. Our method is shown to produce from the raw aggregates pathways with higher fidelity than sampling a Markov chain model of the aggregate data, even preserving the same length of the original pathways. In particular, we observe a relative increase in average cosine similarity of up to 52% with respect to the true pathways compared with aggregate Markov chain sampling.