Interpretability on Clinical Analysis from Pattern Disentanglement Insight

Anonymous

Abstract: Diagnosis of a clinical condition can help medical professionals save time in clinical decision-making and prevent overlooking risks. Therefore we explore the problem of clinical text interpretability using free-text medical notes recorded in electronic health records (EHR). MIMIC-III is a de-identified EHR database containing observations from over 40,000 patients in critical care units. Since medical notes are free-text, existing machine learning models may have ineffective interpretability; however, interpretability is often desirable for clinical diagnosis. Hence, in this paper, we propose a text mining and pattern discovery solution to discover strong association patterns from patient discharge summaries and the code of international classification of diseases (ICD9 code). The proposed approach offers a straightforward interpretation of the underlying relation of patient characteristics in an unsupervised machine learning setting. The clustering results outperform the baseline clustering algorithm and are comparable to baseline supervised methods.

Paper Type: long