Deep Dynamic Patient Similarity Analysis: Model Development and Validation in ICU

• DOI: 10.1016/j.cmpb.2022.107033
• Publication Date: 2022-07-20T15:50:41Z
• Authors: Zhaohong Sun; Xudong Lu; Huilong Duan; Haomin Li
• Keywords: Similarity (geometry); Similarity measure; Personalized Medicine; Baseline (sea)

Personalized medicine requires the patient similarity analysis for providing specific treatments tailed each patient. However, in personalized clinical scenarios encounters challenges, which are twofold. First, heterogeneous and multi-type data usually recorded to Electronic Health Records (EHRs) during course of admissions, makes it difficult measure similarity. Second, disease progression manifests diverse states at different times, brings sequential complexity dynamically retrieve similar patients' sequences.To overcome above-mentioned we propose a novel dynamic model based on deep learning. Specifically, proposed embeds into hidden representations with specially designed embedding attention module. Thereafter, retrieves sequences these manner. More importantly, adopt two tasks, i.e., diagnosis prediction medication recommendation, validate effectiveness model. It is worth noticing that integrates drug-drug interaction (DDI) knowledge graph recommendation task reduce adverse reactions caused by combinational treatments, such more rational strategy can be realized. We evaluate our using critical care database MIMIC-III, includes 5,430 patients covering 14,096 visits.The outperforms several state-of-the-art methods. For prediction, average PR-AUC score reaches 0.6200, significantly higher than baseline models (0.2497∼0.5407). Meanwhile, 0.6682 (Jaccard: 0.4070; F1: 0.5672; Recall: 0.7832) whereas K-nearest only reach 0.3805 0.3911; 0.5465; 0.5705). In addition, achieves lower DDI rate.We model, implemented decision support system tasks including surgical procedure selection, etc. Also, serves as an explainable protocol practice thanks its analogy real reasoning where doctor diagnoses diseases prescribes medications according previous cured empirically.