The identification of latent risk factors that can induce to health risks or an abnormal status is an important task in healthcare data analyses.In recent years,health analyses based on neural network models have been...The identification of latent risk factors that can induce to health risks or an abnormal status is an important task in healthcare data analyses.In recent years,health analyses based on neural network models have been applied widely.However,such analysis processes are blackbox and the results lack explainability.Some approaches by constructing a domain model may tackle these issues.However,domain knowledge from an expert is required.In this study,we propose an exploratory and interpretable approach to estimating latent health risk factors without relying on domain knowledge,in which feature selection and causal discovery are used to construct a domain model for uncovering complex relationships in health and medical data.An evaluation experiment conducted on two datasets by comparing the proposed approach with four baselines demonstrated that the proposed approach outperformed the baselines in terms of model fitness.Furthermore,the number of model parameters in our method was smaller than that in the baselines,which reduced model complexity.Moreover,the analysis process of the proposed approach was visible and explainable,which improved the interpretability of the analysis processes.展开更多
基金supported in part by 2022–2024 Masaru Ibuka Foundation Research Project on Oriental Medicine,2020–2025 JSPS A3 Foresight Program(No.JPJSA3F20200001)2022–2024 Japan National Initiative Promotion for Digital Rural City,2022–2024 JST SPRING(No.JPMJSP2128)+1 种基金2023 and 2024 Waseda University Grants for Special Research Projects(Nos.2023C-216 and 2024C-223)2023–2024 Waseda University Advanced Research Center Project for Regional Cooperation Support,and 2023–2024 Waseda University Advanced Research Center for Human Sciences Project(No.BA080Z000300).
文摘The identification of latent risk factors that can induce to health risks or an abnormal status is an important task in healthcare data analyses.In recent years,health analyses based on neural network models have been applied widely.However,such analysis processes are blackbox and the results lack explainability.Some approaches by constructing a domain model may tackle these issues.However,domain knowledge from an expert is required.In this study,we propose an exploratory and interpretable approach to estimating latent health risk factors without relying on domain knowledge,in which feature selection and causal discovery are used to construct a domain model for uncovering complex relationships in health and medical data.An evaluation experiment conducted on two datasets by comparing the proposed approach with four baselines demonstrated that the proposed approach outperformed the baselines in terms of model fitness.Furthermore,the number of model parameters in our method was smaller than that in the baselines,which reduced model complexity.Moreover,the analysis process of the proposed approach was visible and explainable,which improved the interpretability of the analysis processes.
