Medical Data Privacy Protection: Machine Learning-Driven Secure Matching Technology for Patient Records

This article focuses on how to use machine learning to achieve secure matching of cross-institutional medical records while protecting patient privacy. The study compares various supervised learning models and sampling strategies, and conducts performance analysis and trade-off studies based on real medical data. Key findings include that moderately complex models (such as single-layer neural networks) perform best, and feature representation and class imbalance handling have significant impacts on performance, providing practical guidance for medical data integration and privacy protection.

Privacy Dilemmas in Medical Data Integration

In today's healthcare system, patient health information is scattered across EHR systems of multiple institutions. Integrating scattered data is crucial for continuous medical services and medical research. However, traditional record matching relies on personally identifiable information (PII), which raises privacy concerns, security risks, and HIPAA compliance challenges.

Overview of Privacy-Preserving Record Linkage (PPRL) Technology

PPRL achieves secure matching through transformed or encoded representations instead of raw identifiers, with the core being to identify different records of the same patient without exposing sensitive information. This study, conducted in collaboration between the Regenstrief Institute and Indiana University's Luddy School, uses a real dataset containing 10,000 pairs of labeled records to explore the application of ML in PPRL.

Data Representation and Feature Engineering

Binary consistency features are used to represent record pairs: a feature value of 1 indicates that the fields are consistent, and 0 indicates inconsistency. Features are generated from transformed medical identifiers to encode the degree of matching. The dataset has class imbalance (matching pairs are far fewer than non-matching pairs).

Class Imbalance Handling Strategies

Four strategies are evaluated:

• Original distribution (baseline)
• Oversampling (duplicate minority class)
• Undersampling (reduce majority class)
• SMOTE (synthesize minority class samples)

Evaluated Machine Learning Models

Six models are compared:

1. Logistic Regression: A baseline interpretable model that provides a reference benchmark.
2. SVM: Uses kernel tricks to handle non-linear data.
3. KNN: Classifies based on similarity.
4. Single-Layer Neural Network (SLNN): Evaluates the effect of limited non-linear learning.
5. Multi-Layer Neural Network (MLNN): Explores the impact of deep architectures.
6. XGBoost: A gradient-boosted ensemble model that learns complex feature interactions.

Evaluation Metrics and Statistical Validation

Accuracy, precision, recall, F1 score, and confusion matrix are used, with a focus on the precision-recall trade-off (balancing false positives/negatives). Statistical significance of performance differences is verified through McNemar's test, paired t-test, and Wilcoxon signed-rank test.

1. Single-Layer Neural Network (SLNN) achieves the highest overall performance; moderate complexity is better than simple linear or complex deep networks.
2. Logistic Regression, SVM, and MLNN have statistically comparable performance; increasing complexity does not necessarily improve results.
3. XGBoost has lower recall, possibly due to sensitivity to class imbalance.
4. KNN has high recall but low precision, with more false positives.
5. Simple interpretable models perform as well as complex architectures; performance differences are more affected by feature representation and imbalance handling than model complexity.

Practical Implications

1. Compliance: PPRL helps medical institutions meet privacy regulations such as HIPAA and achieve data integration.
2. Efficiency: Reduces the risk of sensitive information exposure, which is better than traditional PII matching.
3. Interpretability: Simple models perform well, making them easy to deploy and maintain in practice.

Future Outlook

Explore advanced privacy technologies such as federated learning and differential privacy to improve linkage accuracy under secure conditions.

Privacy-preserving record linkage is an important advancement in medical information technology, balancing data utility and patient privacy. Through systematic model comparisons and rigorous statistical validation, this study provides practical guidance for practitioners. As medical data grows and privacy regulations become stricter, PPRL will play a key role in building a secure and efficient medical data ecosystem.