Heart Sound Abnormality Detection from Multi-Institutional Collaboration: Introducing a Federated Learning Framework

Information Science Artificial Intelligence Medicine Cardiovascular System
Heart Sound Abnormality Multi-Institutional Collaboration Federated Learning Privacy Protection Deep Learning

利用联邦学习检测心音异常的一项多机构合作研究

Academic Background

Cardiovascular diseases (CVDs) have become one of the leading causes of death, particularly within the elderly population, making cardiovascular health a pressing societal concern. Early screening, diagnosis, and prognosis management are crucial for preventing hospitalizations. Heart sound signals carry rich physiological and pathological information, providing advantages such as ease of access, widespread presence, and non-invasive nature for early CVD diagnosis. In recent years, the application of artificial intelligence (AI) in heart sound-assisted diagnosis has garnered widespread attention, with automatic heart sound auscultation technology facilitating the rapid and effective assessment of cardiac conditions. However, existing studies often overlook data security and privacy issues, especially in the context of multi-institutional data collaboration.

Research Origin

This paper was authored by Wanyong Qiu, Chen Quan, and others from notable academic institutions such as Beijing Institute of Technology, The University of Tokyo’s Educational Physiology Laboratory, and Imperial College London. The paper was published in the IEEE Transactions on Biomedical Engineering in 2023.

Research Objectives and Methods

Background Issue

Current AI models typically require a large amount of training data, which may raise privacy concerns. Additionally, due to the inability to freely exchange data between medical institutions, data silo issues arise, hindering the collaborative training of AI models across multiple centers and limiting the development of medical AI models.

Research Methods

To address these issues, this paper proposes an optimized strategy based on Federated Learning (FL), aiming to train models on multi-center heart sound databases without disclosing information. The study primarily applies Horizontal Federated Learning (HFL) and Vertical Federated Learning (VFL), with the former addressing privacy concerns and the latter tackling model interpretability and data scarcity issues.

Specifically, the paper innovates in the following areas:

  1. HFL Model Setup: Utilizing HFL to handle multi-institutional heart sound data, aligning feature spaces and securely aggregating IDs for different participating medical institutions.
  2. VFL Model Setup: Jointly training and evaluating the VFL model across multiple institutions’ data feature spaces to address the issue of missing data labels.
  3. Model Interpretability: Employing Shapley values for interpreting the VFL model, balancing model interpretability and data privacy.

Research Process

Data Description and Preprocessing

Data were obtained from heart sound databases of various medical institutions, such as the MIT heart sound database and Aalborg University heart sound database, with data labeled as normal and abnormal samples. Data preprocessing included signal processing, feature extraction, and data balancing strategies.

HFL Model

Using XGBoost as the base model, a horizontal federated learning framework was established. By aligning the feature space with the feature ID safe aggregation method, the specific steps included:

  1. The federated server generates symmetric keys and distributes them to participating institutions.
  2. Each participating institution encrypts its feature ID set using the public key and sends it to the server.
  3. The federated server uses the private key to convert the encrypted results and shares the aggregated results with the participating institutions.
  4. Each participating institution trains the model locally and sends incremental model updates to the server.

VFL Model

In scenarios where the feature space differs but the sample space remains the same, the VFL model builds a global model through multi-party collaboration. The specific steps included:

  1. Dividing the data into guest and host parties, ensuring a consistent sample space but different feature spaces.
  2. Each party performs feature engineering and model training while preserving privacy.
  3. Following privacy protection protocols, all parties jointly evaluate and infer the model.

Main Results and Summary

HFL Model Results

Under conditions of non-independently and identically distributed (non-iid) data and imbalanced sample sizes, the HFL model performed excellently on the test set. The optimal model parameters acquired after 50 repeated experiments included 30 trees with a depth of 3. Sensitivity reached 62.1%, and specificity reached 72.8%, both higher than the traditional XGBoost model.

VFL Model Results

The VFL model performed well in databases with sufficient sample sizes, such as those from Aalborg University and Dalian University of Technology, approaching the performance of traditional centralized data learning models. However, due to significant differences in the data distribution of the df database, the VFL model’s results were below expectations. The Maximum Mean Discrepancy (MMD) value indicated significant distribution differences between the df database and others.

Model Interpretability

By explaining the VFL model with Shapley values, data privacy was retained. In multi-center heart sound databases, federated features enhanced the influence distribution of the global model features on model output. Feature binning methods protected the guest party’s data privacy, yielding the model’s interpretability and evaluating the fairness of data contribution to the VFL model.

Research Significance

This paper is the first to apply federated learning in real medical scenarios to optimize heart sound models, achieving good classification results while protecting patient privacy. The research outcomes indicate that federated learning performs excellently in detecting abnormal heart sounds, potentially promoting the widespread application of federated intelligent healthcare systems, especially in scenarios with high data privacy requirements.

Research Highlights

  1. Multi-institutional Cooperative Model Training: Achieving federated learning in multi-center medical databases while protecting data privacy.
  2. Model Interpretability: Interpreting the VFL model with the Shapley value method, balancing model interpretability and data privacy.
  3. Practical Application Value: Provides a practical guide for using federated learning in heart sound classification, delivering significant value for the development of intelligent healthcare systems.

Future Research Directions

  1. Data Standardization: Establish standardized heart sound databases to address the impact of data heterogeneity and privacy noise on model performance.
  2. Federated Learning Incentive Mechanism Design: Design reasonable incentive mechanisms based on Shapley values to measure feature contributions, encouraging more participants to join federated learning.
  3. Encryption and Communication Costs: Further explore new encryption schemes such as differential privacy, reduce federated learning’s communication costs, and improve the model’s applicability.