Unraveling multimodality of digital health records by comparing mortality trajectories of diagnoses of diseases from over 12 million patients

Understanding the multimodality of digital health data, including the scope of death records, is essential for adequate data acquisition to build a research framework for the health sciences. In this study, I leveraged the diverse healthcare records of over 12 million patients to reconstitute mortality trajectories that navigate the sequence of disease processes shared among patients from initial presentation through interim conditions that ultimately terminate in fatal outcomes. I conducted a comprehensive analysis of longitudinal discharge records for 10.4 million patients from US hospitals, utilizing the US State Inpatient Data (USSID) including 290,253 records of deaths in clinics. I also scrutinized the cross-sectional records of Korea from the billing reviews, specifically the National Inpatients Set of Korea (NISK), encompassing 2.1 million patients. By tracing the diagnostic timelines of patients diagnosed with significant comorbid diseases (False Discovery Rate (FDR) <0.1), I built mortality trajectories, mapping the temporal progression of disease diagnoses resulting in death. My trajectory model rewired 705 significant mortality trajectories across both datasets (USSID and NISK). The presented mortality trajectories successfully recapitulated established patterns of mortality for each country, while also revealing different trajectories leading to death, influenced by the modality of data. For example, viral hepatitis, a known predisposing feature of liver cancer in Asia, was observed to initiate in younger Koreans. Interestingly, owing to the collection of hospital records, the modeled mortality trajectories derived from the USSID converged towards sepsis. Although a substantial sequence of diagnostic processes is shared between USSID and NISK, the multimodality of these two datasets highlights different diagnoses preceded by fatal outcomes. Unraveling mortality patterns is feasible with an appropriate understanding of the multimodality of digital health data.