Using Network Analysis Modularity to Represent Health Code Systems in Machine Learning Models

Background: A common issue facing machine learning (ML) modeling in healthcare datasets is how to represent high-dimensional code systems e.g. International Classification of Diseases (ICD), Anatomical Therapeutic Chemical classification (ATC), and Diagnosis Related Groups (DRG) codes, given that there is always a trade-off between reducing the number of model dimensions and minimizing information loss.

Objectives: To investigate an approach of representing clinical and administrative healthcare-related code systems using network analysis modularity.

Methods: We used the MIMIC-III dataset, an open-source Intensive Care Unit (ICU) dataset to create a network in which ICD-9 codes are the nodes and the number of patients who had the same diagnoses represent the edges. We then applied network modularity with different resolution thresholds to generate 6 different sets of modules. Each set was used separately to represent ICD codes creating 6 different datasets. These modules represent comorbidities occurring in the dataset. We implemented Logistic Regression (LR), Support Vector Classifier (SVC) with non-linear kernel and Gradient Boosting Machines (GBM) to compare the performance of predicting 90-days ICU readmission between 1) the raw dataset, 2) 6 sets of modules' representations, 3) representing ICD codes using single-level Clinical Classification Software (CCS), and 4) truncated ICD codes to their highest level of hierarchy. The same methodology was applied to encoding DRG codes comparing only one set of modules to the raw dataset. Accuracy, AUC, F1-score were used to evaluate performance.

Results: ICD codes: Different encoding approaches performed almost the same compared to encoding the raw dataset both in LR and GBM (AUC ranged from 80-81% in LR, 83-84% in GBM), while they performed better in SVC (with highest AUC 85% versus 82% in the raw dataset). There was no significant difference in performance between the different modules', CCS, or truncation representations.
DRG codes: All models performed better on the modularity encoded dataset than the raw dataset. GBM model yielded the highest AUC 0.8 (95% CI 0.790 - 0.807).
For both ICD and DRG, training time was significantly reduced using modularity encoding in SVC and GBM models.

Conclusions: This is a proof of concept that using modularity encoding performs well in representing health code systems. Modularity encoding also provides additional advantages such as clinical relevance, usability to encode any code system that can be aggregated to the patient level, enhancing model interpretability, substantially fewer features and faster training time. Furthermore, investigating comorbidity modules will provide additional clinical information to ML models.