[Paper Review] Machine Learning in Precision Medicine to Preserve Privacy via Encryption

William Briguglio, Parisa Moghaddam+3 more|arXiv (Cornell University)|Feb 5, 2021

Cancer Genomics and Diagnostics15 references23 citations

TL;DR

This paper proposes a Machine Learning with Encryption (MLE) framework that enables privacy-preserving cancer prediction using homomorphic encryption on genomic data from the MSK-IMPACT dataset. The framework trains a model on encrypted data, achieving 77.47% accuracy—slightly higher than prior work—while ensuring patient genomic privacy through fully homomorphic encryption, with all code and a cloud deployment service released openly.

ABSTRACT

Precision medicine is an emerging approach for disease treatment and prevention that delivers personalized care to individual patients by considering their genetic makeups, medical histories, environments, and lifestyles. Despite the rapid advancement of precision medicine and its considerable promise, several underlying technological challenges remain unsolved. One such challenge of great importance is the security and privacy of precision health-related data, such as genomic data and electronic health records, which stifle collaboration and hamper the full potential of machine-learning (ML) algorithms. To preserve data privacy while providing ML solutions, this article makes three contributions. First, we propose a generic machine learning with encryption (MLE) framework, which we used to build an ML model that predicts cancer from one of the most recent comprehensive genomics datasets in the field. Second, our framework's prediction accuracy is slightly higher than that of the most recent studies conducted on the same dataset, yet it maintains the privacy of the patients' genomic data. Third, to facilitate the validation, reproduction, and extension of this work, we provide an open-source repository that contains the design and implementation of the framework, all the ML experiments and code, and the final predictive model deployed to a free cloud service.

Motivation & Objective

• To address the critical challenge of privacy leakage in precision medicine when sharing sensitive genomic and electronic health data.
• To develop a machine learning framework that enables model training and inference on encrypted data without exposing raw patient information.
• To improve predictive accuracy for cancer risk using the MSK-IMPACT genomic dataset while preserving data privacy.
• To facilitate reproducibility and community extension by releasing open-source code, system design, and a free cloud-based prediction service.

Proposed method

• The MLE framework integrates homomorphic encryption (HE) using the BFV scheme from Microsoft SEAL to perform computations on encrypted genomic features.
• Feature preprocessing includes standardization, log transformation, and selection via mutual information or chi-squared tests to optimize model performance.
• Linear models (logistic regression and SVM) are selected as HE-friendly classifiers due to their compatibility with homomorphic operations.
• The model computes encrypted linear scores using homomorphic addition and multiplication, then sends only the encrypted scores to the client for final decryption and decision-making.
• The framework supports multiple configurations of feature preprocessing and classifiers, with 2,240 total combinations tested on the MSK-IMPACT dataset.
• A free cloud service is deployed to allow medical practitioners to upload cases and obtain encrypted predictions without exposing raw data.

Experimental results

Research questions

• RQ1 Can a machine learning model be trained and deployed on encrypted genomic data to preserve patient privacy while maintaining high predictive accuracy?
• RQ2 How does the performance of a homomorphic encryption-based model compare to non-encrypted models on the same MSK-IMPACT dataset?
• RQ3 What are the computational trade-offs of different homomorphic encryption parameter settings in terms of accuracy and prediction time?
• RQ4 How can privacy-preserving machine learning be made accessible and reproducible for medical, ML, and software engineering communities?
• RQ5 Can the MLE framework support future integration of more complex models while preserving data privacy?

Key findings

• The MLE framework achieved a prediction accuracy of 77.47% on the MSK-IMPACT dataset, outperforming the most recent published study on the same dataset.
• The model's performance remained stable across different precision levels, with accuracy stabilizing at 77.41% for multipliers from 10^3 to 10^9.
• The best configuration used standardization and chi-squared feature selection with a logistic regression classifier, achieving 77.47% accuracy.
• The framework's encryption time varied significantly with parameter choices, with optimal settings achieving an average prediction time of ~3456 seconds per sample on a standard machine.
• The system achieved 100% agreement in score ranking at precision levels of 10^6 and above, indicating high consistency in relative risk ordering.
• An open-source repository was released, including full code, model deployment, and a free cloud service for medical practitioners to test individual cases securely.