Work

Using Artificial Intelligence to Discern Effective Clinical Practice Patterns using Electronic Health Record Data

Year: 2026
Type: dissertation
Study design: Observational Study
Abstract: Background: Type 2 diabetes mellitus (T2DM) is a complex chronic condition with considerable variability in inpatient management and outcomes. Despite widespread use of electronic health records (EHRs), significant variation in clinical practices remains, which may lead to suboptimal results. Artificial intelligence and machine learning (AI/ML) present opportunities to uncover hidden care patterns within EHR data that could support more consistent, evidence-based decision-making. Purpose: The purpose of this study was to assess the feasibility of using artificial intelligence/machine learning (AI/ML) models with electronic health record (EHR) data and to identify clinical practice patterns (care clusters) associated with effective healthcare outcomes among inpatients with type 2 diabetes mellitus (T2DM). Methods: A two-phase quantitative descriptive design was employed. Phase 1 examined data accessibility, completeness, and the feasibility of identifying latent order clusters using data exploration tools and unsupervised machine learning techniques, including latent Dirichlet allocation (LDA), non-negative matrix factorization (NMF), and latent semantic analysis (LSA). Phase 2 investigated the relationships between the identified care patterns and outcomes through generalized linear models and penalized regression methods such as LASSO and Elastic Net. Data were collected from the MIMIC-IV database, comprising 4,601 admissions that met the inclusion criteria. The primary outcomes measured included length of stay (LOS), time in range (TIR), hypoglycemia rate, and mean glucose. Results: Feasibility was confirmed through successful data extraction, preprocessing, and model implementation. Unsupervised models identified clinically interpretable care-pattern phenotypes, with convergence across different modeling approaches supporting robustness. In Phase 2, specific orders and care patterns were linked to outcomes. Insulin use, steroid use, and increased monitoring were consistently associated with poorer glycemic outcomes and longer LOS, whereas structured care elements such as insulin glargine, diet orders, and diabetes consults showed protective or stabilizing effects. Glycemic outcomes exhibited nuanced patterns, including higher hypoglycemia risk with intensive treatment and improved TIR with structured care processes. Penalized regression models demonstrated stable feature selection, supporting the identification of clinically meaningful predictors. Conclusions: Findings indicate the feasibility of using AI/ML to identify latent care patterns from EHR order data and to link these patterns to clinical outcomes in inpatient T2DM populations. Findings suggest that EHR-derived order clusters can serve as proxies for clinical practice patterns and may inform future development of AI-enabled clinical decision support systems. This work advances the application of the Data to Wisdom continuum and Technology-Enhanced Wisdom framework by illustrating how AI/ML can transform complex clinical data into actionable knowledge to support nursing and interdisciplinary decision-making and subsequent wisdom. Implications: AI/ML-driven identification of care patterns has the potential to reduce unwarranted variation, enhance clinical decision-making, and improve patient outcomes. Further research is needed to validate these findings and translate them into real-time clinical decision support tools.
Author: Vyas, Pankaj
Language: English
Open Access status: green
OpenAlex ID: W7164730740
Cites: 96
Cited by: 1
Related to: 10