Abstract
Hospital readmission within 30 days is a critical indicator of healthcare quality and represents a significant financial burden. This study develops a predictive framework for readmission risk among diabetic patients using a large-scale dataset of 100,244 records. By evaluating Random Forest (RF), Naive Bayes (NB), and Tree Ensemble (ET) classifiers within the KNIME platform, the research identifies Random Forest as the superior model. Key findings highlight the importance of secondary diagnoses and prior inpatient history in predicting risk, offering a low-code pathway for clinical implementation.
Introduction
For patients with diabetes mellitus (DM), 30-day readmission rates are notably high, ranging from 14.4% to 22.7%. These rates are driven by the complexity of glycemic control and various comorbidities. Predicting which patients are at high risk before discharge is challenging due to the inherent "class imbalance" in clinical data—where the majority of patients are not readmitted. This study addresses this gap by building robust machine learning models to identify high-risk individuals, enabling timely clinical interventions.
Methodology
The researchers employed the Konstanz Information Miner (KNIME) platform to compare three distinct machine learning approaches:
- Random Forest (RF): An ensemble method using multiple decision trees.
- Naive Bayes (NB): A probabilistic model based on feature independence.
- Tree Ensemble (ET): A method combining multiple decision trees to improve generalization.
To address the fact that only ~11% of the population experienced 30-day readmission, the study tested two balancing strategies: SMOTE (oversampling) and Equal Size Sampling (down-sampling). Feature engineering was also critical, reducing hundreds of ICD-9 codes into 16 consolidated diagnostic categories.
Dataset Characteristics
The study utilized 100,244 records from the Health Facts Database (Cerner Corporation, US) spanning 1999 to 2008. After preprocessing, 23 modeling risk factors were retained from an initial set of 55 attributes.
As shown in Figure 1, the dataset is predominantly Caucasian (77.08%) and slightly skewed toward female patients (53.79%).
Key Insights and Results
The Random Forest (RF) model consistently outperformed other classifiers. Using a down-sampling strategy, RF achieved an AUC of 0.661 for 30-day readmission and an AUC of 0.686 for general future readmission.
Predictor Importance
A significant finding was that the secondary diagnosis (diag_2) was a stronger predictor than the primary diagnosis. This suggests that the overall clinical complexity of a diabetic patient is more indicative of readmission risk than the acute reason for the initial visit.
The model identified a "high-risk" profile: elderly patients with frequent emergency department visits and initial hospital stays exceeding five days (which correlates with an 87% higher risk compared to stays of two days or less). Figure 3 illustrates the distribution of medication use, another key clinical factor integrated into the model.
The top five influential features were:
- Number of inpatient admissions
- Age
- Number of emergencies
- Secondary diagnosis (
diag_2) - Sex
Conclusion
This research demonstrates that machine learning, specifically Random Forest models, can effectively flag diabetic patients at risk for short-term readmission. By utilizing accessible platforms like KNIME, these insights can be integrated into clinical workflows without requiring deep programming expertise. Future research will focus on incorporating more recent data and refining the impact of specific comorbidities on readmission timelines.
Summary updated