Federated learning with multi-cohort real-world data for predicting the progression from mild cognitive impairment to Alzheimer's disease

Abstract

Introduction: Leveraging routinely collected electronic health records (EHRs) from multiple health-care institutions, this approach aims to assess the feasibility of using federated learning (FL) to predict the progression from mild cognitive impairment (MCI) to Alzheimer's disease (AD).

Methods: We analyzed EHR data from the OneFlorida+ consortium, simulating six sites, and used a long short-term memory (LSTM) model with a federated averaging (FedAvg) algorithm. A personalized FL approach was used to address between-site heterogeneity. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and feature importance techniques.

Results: Of 44,899 MCI patients, 6391 progressed to AD. FL models achieved a 6% improvement in AUC compared to local models. Key predictive features included body mass index, vitamin B12, blood pressure, and others.

Discussion: FL showed promise in predicting AD progression by integrating heterogeneous data across multiple institutions while preserving privacy. Despite limitations, it offers potential for future clinical applications.

Highlights: We applied long short-term memory and federated learning (FL) to predict mild cognitive impairment to Alzheimer's disease progression using electronic health record data from multiple institutions. FL improved prediction performance, with a 6% increase in area under the receiver operating characteristic curve compared to local models. We identified key predictive features, such as body mass index, vitamin B12, and blood pressure. FL shows effectiveness in handling data heterogeneity across multiple sites while ensuring data privacy. Personalized and pooled FL models generally performed better than global and local models.

Keywords: Alzheimer's disease; federated learning; long short‐term memory; mild cognitive impairment.

FIGURE 1

Federated learning scenario: simulation of data‐restricted areas in North Florida, Central Florida, South Florida, Georgia, Alabama, and Other Linked Data Regions.

FIGURE 2

Local model and federated learning system. Local LSTM prediction model (A). This architecture includes two key parts: input data and model. The input data part contains temporal trajectories of EHRs for 𝑛 patients, represented as 𝑥^𝑡 _𝑖, where 𝑖 ∈ {1, 2, …, 𝑛} and 𝑡 represents time points. The model part features a combination of LSTM layers to encode temporal trajectories and an MLP predictor to forecast whether patients will convert to AD based on the EHRs. System architecture of federated learning (B). The federated learning process starts with the generation of an initial global model on the federation server. Each participating site receives this model and trains it locally using its private data. After a specified number of training epochs, each site sends its local model weights back to the server. The server then applies the FedAvg algorithm to synthesize these weights into an updated global model, which is subsequently redistributed to each site, replacing the initial weights. AD, Alzheimer's disease; EHRs, electronic health records; FedAvg, federated averaging; LSTM, long short‐term memory; MLP, multilayer perceptron.

FIGURE 3

The average Shapley additive explanation (SHAP) values of various health indicators and disease conditions on model outputs in different regions (Alabama and others) and under two conditions (within 24 months and unrestricted). The SHAP values indicate the relative importance and direction of influence of each feature in model decisions. The left subfigures represent the mean of numerical values, while the right subfigures have red portions representing high impacts and blue portions representing low impacts, with the specific values demonstrating the magnitude of the average impacts. B12, vitamin B12; BMI, body mass index; BP, blood pressure; HDL, high‐density lipoprotein.

FIGURE 4

Comparison of feature temporal importances in global and personalized models across different regions: (A) Alabama and (B) Other Linked Data Regions. Each bar represents the difference of AUC by disrupting a feature of data, with the global model shown in blue and the personalized model in orange. AUC, area under the receiver operating characteristic curve; B12, vitamin B12; BMI, body mass index; BP, blood pressure; HDL, high‐density lipoprotein.

FIGURE 5

Tendency of key biomarkers in Alabama and Other Linked Data Regions. This figure illustrates the trends over time for six biomarkers: B12, BMI, blood pressure (systolic and diastolic), cholesterol, and HDL. Data were collected every 3 months, with the mean values plotted. The solid red line represents the mean for patients who progressed from MCI to AD, and the solid blue line represents the mean for patients who remained stable at the MCI stage, while the shaded areas indicate one standard deviation. AD, Alzheimer's disease; B12, vitamin B12; BMI, body mass index; BP, blood pressure; HDL, high‐density lipoprotein; MCI, mild cognitive impairment.