Abstract

This study introduces a novel approach to Electronic Health Record (EHR) analysis, extending the use of phenotyping with machine learning (ML) models to enhance the recognition and treatment of anemia. It first examines the healthcare scenario in India and suggests potential improvements through data-driven personalized care. Using the MIMIC-III dataset, the research involves extensive data preprocessing and analysis to uncover key insights into anemia's prevalence, gender distribution, comorbidities, and Intensive Care Unit (ICU) stays. Partitioning clustering algorithms like K-Means, K-medoids, Fuzzy C-means, and hierarchical clustering algorithms such as Agglomerative Clustering, DIANA, and HDBSCAN are used to identify groups of patients with similar medical profiles. The distance metrics employed are Levenshtein and Euclidean distances combined with TF-IDF Vectorization. The effectiveness of these algorithms is evaluated based on Length of Stay (LoS) estimation, a critical parameter in EHR studies. To predict a new patient's LoS, the patient is at first classified into an existing cluster, which shows the highest support to the patient’s clinical activities. A decision tree regressor is then trained using data from the selected cluster to predict the new patient's LoS, significantly improving predictive accuracy and reliability. Notably, the HDBSCAN algorithm, applied to the Tf-Idf Vectorizer object, achieves a 50.82% reduction in Root Mean Squared Error (RMSE) compared to baseline model. The novelty of this study lies in proposing an efficient approach for EHR analysis, specifically for predicting ICU patients' LoS, and identifying the most effective clustering algorithm to improve healthcare delivery for anemic patients in healthcare scenario of India.

Statistics