Machine Learning Versus Self-Supervised Transfer Learning for 30-Day Readmission Prediction in Diabetic Patients

Technical Summary: Machine Learning Versus Self-Supervised Transfer Learning for 30-Day Readmission Prediction in Diabetic Patients

Problem Statement
Hospital readmission within 30 days of discharge is a critical quality indicator and a major driver of healthcare costs, particularly for patients with diabetes. While electronic health records (EHRs) contain vast amounts of structured data, predicting readmission risk using only administrative and coded clinical features remains challenging. The study addresses two primary gaps: first, the lack of consensus on whether deep learning (specifically self-supervised transfer learning) offers a genuine advantage over established tree-based ensembles for tabular clinical data; and second, the methodological pitfalls often associated with class imbalance handling and decision threshold selection in this domain. The authors aim to establish realistic performance expectations for models built solely on administrative data and to demonstrate the critical importance of threshold selection over model architecture.

Methodology
The study utilized the UCI Diabetes 130-US Hospitals dataset, comprising 101,766 inpatient encounters from 1999–2008. After excluding encounters ending in death or hospice, 99,340 encounters were stratified into 70/15/15 training, validation, and test partitions, maintaining an 11.4% readmission prevalence.

The methodology compared six predictive models under a unified, leakage-controlled protocol:

1. Classical Classifiers: Logistic Regression, Random Forest, XGBoost, and LightGBM.
2. Neural Networks: Two Multilayer Perceptrons (MLPs). One was trained from scratch with random initialization. The second utilized self-supervised transfer learning, where a Denoising Autoencoder (DAE) was pretrained on the full unlabeled feature matrix (using Gaussian noise and random masking) to learn robust representations. These encoder weights were then transferred to initialize the MLP's first hidden layer before fine-tuning on labeled data.

Key Methodological Adaptations for Tabular Data:

• Transfer Learning: Mapped to DAE pretraining on the feature matrix rather than external image/text corpora.
• Data Augmentation: Mapped to the Synthetic Minority Over-sampling Technique (SMOTE), applied strictly within cross-validation folds to prevent data leakage.
• Pretrained Family: Gradient-boosted trees (XGBoost/LightGBM) were treated as the "pretrained" baseline for structured data, consistent with literature suggesting they are state-of-the-art for tabular tasks.

A critical methodological contribution was the identification and correction of a data leakage error. An initial hyperparameter search applied SMOTE to the entire training set before cross-validation, yielding an implausible AUROC of 0.957. This was corrected by integrating SMOTE into an imblearn pipeline to ensure resampling occurred only within the training fold of each cross-validation iteration, resulting in a corrected AUROC of 0.661.

Key Results

• Model Performance: Tree-based ensembles significantly outperformed neural networks. XGBoost achieved the highest validation AUROC (0.672), followed by LightGBM (0.669) and Random Forest (0.654). Both MLP variants (from-scratch and transfer-learned) lagged behind with AUROCs of 0.585 and 0.582, respectively. The self-supervised pretraining did not improve the MLP's performance; if anything, it was marginally worse.
• Hyperparameter Tuning: After correcting for leakage, tuned XGBoost achieved a test AUROC of 0.675 and an Area Under the Precision-Recall Curve (AUPRC) of 0.244.
• Threshold Sensitivity: The study highlighted a stark contrast in clinical utility based on the decision threshold.
  • At the default threshold of 0.5, the model achieved 88.6% accuracy but a sensitivity of only 1.4%, rendering it useless for screening.
  • Using a threshold optimized via Youden's J statistic (0.128) on the validation set increased sensitivity to 58.0% but reduced precision to 18.5%.
• Predictors: The most predictive features included circulatory system diagnoses, prior inpatient visits, and missing Hemoglobin A1c results, aligning with clinical intuition.

Significance and Claims
The paper claims three primary contributions:

1. Performance Ceiling: The results converge with two independent 2025–2026 studies using the same dataset (reporting AUROCs of 0.667 and 0.664–0.688). This convergence suggests a practical discrimination ceiling of 0.66–0.69 AUROC for 30-day readmission prediction using only administrative and coded clinical features. The authors argue that further gains require richer data (e.g., lab trends, unstructured text, social determinants) rather than more complex classifiers.
2. Inefficacy of Transfer Learning on Tabular Data: The study provides empirical evidence that, for medium-sized tabular clinical datasets, self-supervised pretraining via a simple denoising autoencoder does not offer an advantage over random initialization and is inferior to gradient-boosted trees.
3. Methodological Rigor and Threshold Selection: The paper emphasizes that decision threshold selection is a "first-class methodological choice." It demonstrates that a model with moderate discrimination can be clinically inert (at 0.5 threshold) or a usable screening tool (at an optimized threshold), depending entirely on the chosen operating point. It also serves as a cautionary tale regarding SMOTE-induced data leakage in cross-validation.

Conclusion
The study concludes that for readmission risk stratification using administrative EHR data, well-tuned gradient-boosted trees (specifically XGBoost) remain the superior choice over neural networks, even when the latter are enhanced with self-supervised pretraining. The findings advocate for realistic expectations regarding the predictive power of administrative data alone and stress that the deployment of such models requires explicit, cost-aware threshold selection rather than reliance on default settings.