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.

1. BACKGROUND

Alzheimer's disease (AD) is a significant public health concern as a progressive neurodegenerative disorder affecting millions worldwide, leading to a gradual decline in cognitive functions, including memory and problem‐solving abilities. ¹ , ² Mild cognitive impairment (MCI) serves as a transitional state between normal cognitive function and dementia, characterized by diverse cognitive profiles and clinical progression patterns, making it challenging to predict outcomes and progression in MCI patients. ³ , ⁴ An estimated 10% to 15% of individuals with MCI develop dementia annually, with approximately one third of those with MCI due to AD progressing to dementia within 5 years. ⁵ , ⁶ Identifying high‐risk MCI patients is key for early prevention, trial stratification, and personalized management. ⁷

Various data modalities have been explored to predict the progression from MCI to AD, including biomarkers, ⁸ neuroimaging, ⁹ , ¹⁰ neuropsychological tests, clinical measures, ¹¹ genetic data, ¹² digital health data, demographic and lifestyle variables, as well as their combinations. ⁷ , ¹¹ , ¹³ However, these data sources are often difficult to obtain and not routinely collected, which may lead to delayed disease awareness and diagnosis.

In contrast, routinely collected clinical data from electronic health records (EHRs) reflect real‐world evidence and contain a wealth of information that can be leveraged to predict AD. EHRs include longitudinal patient data that help track cognitive decline. Over the past decades, several studies have been conducted to predict AD using EHRs. Li et al. ¹⁴ used machine learning (ML) with EHRs to identify factors predicting early AD onset. Park et al. ¹⁵ explored using ML and large‐scale health data to predict AD. Moreover, researchers have also focused on predicting the transition from MCI to AD. Fouladvand et al. ¹⁶ used ML algorithms incorporating structured EHRs and unstructured notes to forecast the conversion from MCI to AD. Pang et al. ¹⁷ used National Alzheimer's Coordinating Center (NACC) ¹⁸ data to predict progression from normal cognition to MCI and AD. Despite promising results, these models are often tested on region‐specific datasets. This limitation can hinder the generalizability of the results, as the models might not perform consistently when applied to data from diverse demographic or geographic backgrounds. Thus, the findings may not be generally applicable across varied health‐care settings.

Although EHRs are relatively easy to collect, coordinating EHRs across diverse health‐care institutions presents major challenges, compounded by the need to protect health information and concerns over patient privacy and data security. These challenge hinder traditional centralized data analysis methods. Federated learning (FL) offers a solution, allowing private, secure model training across sites without sharing raw data. ¹⁹ , ²⁰ , ²¹ FL has been successfully applied to various tasks, including patient similar learning, ²² , ²³ phenotyping, ²⁴ , ²⁵ mortality prediction, ²⁶ , ²⁷ hospitalization prediction, ¹⁹ disease prediction, ²⁸ , ²⁹ and across different disease domains such as COVID‐19, ²⁷ , ³⁰ cardiovascular disease, ³¹ , ³² cancer, ³³ , ³⁴ kidney diseases, ³⁵ , ³⁶ diabetes, ³⁷ and AD, ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ among others. These studies confirm FL's effectiveness in enhancing prediction and detection. However, there is limited research on FL's use in predicting MCI to AD progression with EHRs from various institutions.

In this paper, we used EHRs from OneFlorida+ and partitioned the data by location to collaboratively train a model for predicting the progression from MCI to AD. We used a long short‐term memory (LSTM) model ⁴⁷ to capture the temporal dependencies in the data. To account for heterogeneity in data distributions among different sites, we adopted a personalized FL approach. This involved fine‐tuning a global model with local data to capture site‐specific traits. To validate whether the model exploits the heterogeneity of data distribution across sites, we assessed the importance of different features using Shapley additive explanation (SHAP) values. ⁴⁸ Additionally, we also explored the permutation feature importance (PFI) ⁴⁹ by comparing the effect of each feature on the model's area under the receiver operating characteristic curve (AUC). By combining LSTM and FL, our approach aims to provide a robust and privacy‐preserving method for predicting the progression of MCI to AD across diverse health‐care settings.

2. METHODS

2.1. Data collection and processing

2.1.1. Data source and study population

Our study used a large repository of EHRs from the OneFlorida+ Clinical Research Consortium. ⁵⁰ This consortium is a member of the National Patient‐Centered Clinical Research Network (PCORnet) and represents a collaborative effort among 14 health organizations. These organizations span academic health centers, community health systems, and clinics, collectively covering ≈ 20 million patients across Florida (≈ 18 million), Georgia (≈ 1 million), and Alabama (≈ 2.1 million). The OneFlorida+ dataset adheres to the PCORnet Common Data Model (CDM), encompassing comprehensive patient information, such as demographics, enrollment status, vital signs, conditions, encounters, diagnoses, procedures, prescriptions, dispensing, and laboratory results.

To simulate a FL environment and account for heterogeneity between hospitals due to geographic differences, we partitioned the dataset into six sites based on state and geographic areas, as shown in Figure 1: in Florida, the data were further segmented into four regions: North Florida (e.g., University of Florida Health), Central Florida (e.g., Tampa, Orlando), South Florida (e.g., Miami), and an additional site comprising linked data with unspecified location information in Florida and two other states including Georgia (i.e., Emory University) and Alabama (i.e., University of Alabama at Birmingham).

FIGURE 1.

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

RESEARCH‐IN‐CONTEXT

1. Systematic review: We reviewed the existing literature using traditional sources, such as PubMed, alongside conference abstracts, presentations, and other relevant materials. While the use of electronic health record (EHR) data to predict the transition from mild cognitive impairment (MCI) to Alzheimer's disease (AD) is still an emerging area of research, recent studies have explored the clinical aspects of this approach, with all pertinent references accurately cited.

2. Interpretation: Our study demonstrates the potential of federated learning (FL) using real‐world EHR data to predict the progression from MCI to AD, despite several limitations. FL's ability to integrate data across institutions and adapt to data heterogeneity offers promising results for future clinical applications.

3. Future directions: To improve the robustness of FL models, future research should focus on enhancing data quality, addressing inconsistencies in diagnostic coding and incorporating additional data modalities, such as clinical notes. Expanding the dataset to include diverse populations from more sites across various geographic areas and using more advanced modeling techniques will increase generalizability. Additionally, integrating FL‐based prediction tools into clinical workflows, such as shared decision‐making apps, could help personalize risk assessments and support patient–physician decisions.

The following criteria were used to determine research participants’ eligibility: (1) an MCI diagnosis made after January 2012, (2) being ≥ 50 years of age at the time of the first MCI diagnosis, and (3) having at least 1 year of data available from the first MCI diagnosis. The study excluded patients who had been diagnosed with AD prior to the MCI diagnosis. Instances of MCI were identified using International Classification of Diseases (ICD) codes 331.83 and 294.9 from ICD‐9, and G31.84 and F09 from ICD‐10. AD was identified using ICD codes 331.0 (ICD‐9) and G30.* (ICD‐10). The characteristics of the study cohort are shown in Table 1.

TABLE 1.

Characteristics of the study cohort.

Characteristics	Total (N = 44,899)	North Florida (N = 6765)	Central Florida (N = 5888)	South Florida (N = 3129)	Alabama (N = 11,239)	Georgia (N = 3151)	Other (N = 14,727)
Age, mean ± std	71.5 ± 10.3	72.4 ± 9.8	72.5 ± 10.7	70.7 ± 10.0	71.5 ± 10.5	73.4 ± 9.5	70.5 ± 10.4
Female, N (%)	23,708 (52.8)	3226 (47.7)	3050 (51.8)	1558 (49.8)	6182 (55.0)	1709 (54.2)	7983 (54.2)
Hispanic, N (%)
Hispanic	4189 (9.3)	223 (3.2)	515 (8.7)	1136 (36.3)	57 (0.5)	50 (1.6)	2208 (15.0)
Not Hispanic	36,114 (80.4)	6429 (95.0)	4324 (73.4)	1837 (58.7)	10,274 (91.4)	1879 (59.6)	11,371 (77.2)
Unknown	4596 (10.2)	113 (1.7)	1049 (17.8)	156 (5.0)	908 (8.1)	1222 (38.8)	1148 (7.8)
Race, N (%)
American Indian or Alaska Native	75 (0.2)	9 (0.1)	21 (0.4)	3 (0.1)	11 (0.1)	4 (0.1)	27 (0.2)
Asian	516 (1.1)	0 (0.0)	77 (1.3)	32 (1.0)	181 (1.6)	46 (1.5)	180 (1.2)
Black or African American	6813 (15.1)	613 (9.1)	487 (8.3)	178 (5.7)	2243 (20.0)	960 (30.5)	2332 (15.8)
Native Hawaiian or Other Pacific Islander	25 (0.1)	1 (0.0)	3 (0.1)	1 (0.0)	1 (0.0)	1 (0.0)	18 (0.1)
White	34,129 (76.0)	5773 (85.3)	4523 (76.8)	2690 (86.0)	8143 (72.5)	1739 (55.2)	11,261 (76.5)
Multiple race	91 (0.2)	23 (0.3)	21 (0.4)	0 (0.0)	0 (0.0)	0 (0.0)	47 (0.3)
Unknown	3250 (7.2)	346 (5.1)	756 (12.8)	225 (7.2)	660 (5.9)	401 (12.7)	862 (5.9)
Conversion and encounters, N (%)
# Convert from MCI to AD	6391 (14.2)	846 (12.5)	721 (12.2)	286 (9.1)	1847 (16.4)	380 (12.1)	2311 (15.7)
# Average convert days (days)	723.6	716.5	602.5	671	688.5	225	873.6
# Median convert days (days)	490.5	564	388	479	425	193	671
# Do not convert	38,508 (85.8)	5919 (87.5)	5167 (87.8)	2843 (90.9)	9392 (83.6)	2771 (87.9)	12,416 (84.3)
# Average encounters	139.6	111.6	37.1	154.9	57.3	54.9	271
# Average encounter length (days)	2099.5	1941.1	1656.2	2232.4	2089.7	470.9	2677.2
Related disease, N (%)
Other cerebral degenerations	29,133 (64.9)	3715 (54.9)	4099 (69.6)	1994 (63.7)	7031 (62.6)	2855 (90.6)	9439 (64.1)
Essential hypertension	24,942 (55.6)	3560 (52.6)	3435 (58.3)	706 (22.6)	6179 (55.0)	1254 (39.8)	9808 (66.6)
Hyperlipidemia	18,999 (42.3)	2786 (41.2)	2696 (45.8)	397 (12.7)	4685 (41.7)	636 (20.2)	7799 (53.0)
Neurological disorders	16,931 (37.7)	3232 (47.8)	1645 (27.9)	342 (10.9)	5033 (44.8)	497 (15.8)	6182 (42.0)
Hypertension	15,252 (34.0)	2384 (35.2)	2097 (35.6)	–	4235 (37.7)	–	6536 (44.4)
Disorders of lipoid metabolism	12,088 (26.9)	1912 (28.3)	1601 (27.2)	–	3362 (29.9)	–	5213 (35.4)
Malaise and fatigue	11,983 (26.7)	2012 (29.7)	1327 (22.5)	274 (8.8)	3019 (26.9)	308 (9.8)	5043 (34.2)
Memory loss	11,794 (26.3)	1881 (27.8)	1652 (28.1)	586 (18.7)	3137 (27.9)	629 (20.0)	3909 (26.5)
Pain in joint	11,649 (25.9)	1902 (28.1)	1002 (17.0)	377 (12.1)	3093 (27.5)	302 (9.6)	4973 (33.8)
GERD	11,615 (25.9)	1704 (25.2)	1321 (22.4)	266 (8.5)	2985 (26.6)	306 (9.7)	5,033 (34.2)
Related drugs, N (%)
Other analgesics and antipyretics	15,085 (33.6)	2562 (37.9)	2174 (36.9)	373 (11.9)	3270 (29.1)	233 (7.4)	6473 (44.0)
Antidepressants	13,325 (29.7)	2233 (33.0)	1072 (18.2)	862 (27.6)	4047 (36.0)	506 (16.1)	4605 (31.3)
Stemmatological preparations	13,180 (29.4)	2447 (36.2)	1571 (26.7)	489 (15.6)	2966 (26.4)	223 (7.1)	5484 (37.2)
Opioids	12,594 (28.1)	2219 (32.8)	1582 (26.9)	359 (11.5)	2716 (24.2)	222 (7.1)	5496 (37.3)
Lipid modifying agents, plain	11,969 (26.7)	2299 (34.0)	1156 (19.6)	445 (14.2)	2964 (26.4)	427 (13.6)	4678 (31.8)
I.V. solution additives	11,947 (26.6)	2702 (39.9)	1491 (25.3)	287 (9.2)	2444 (21.8)	91 (2.9)	4932 (33.5)
Antithrombotic agents	11,611 (25.9)	2382 (35.2)	1688 (28.7)	201 (6.4)	2145 (19.1)	222 (7.1)	4973 (33.8)
Decongestants and other nasal preparations for topical use	11,460 (25.5)	2374 (35.1)	1154 (20.0)	605 (19.3)	2555 (22.7)	188 (6.0)	4584 (31.1)
Agents for treatment of hemorrhoids and anal fissures for topical use	11,408 (25.4)	2419 (35.8)	1241 (21.1)	583 (18.6)	2260 (20.1)	173 (5.5)	4732 (32.1)
Anti‐infectives	11,298 (25.2)	1958 (28.9)	1213 (20.6)	569 (18.2)	2660 (23.7)	104 (3.3)	4794 (32.6)

Abbreviations: AD, Alzheimer's disease; GERD, gastroesophageal reflux disease; I.V., intravenous; MCI, mild cognitive impairment.

2.1.2. Temporal trajectory composition and processing

As mentioned previously, EHRs encompass a variety of data points such as medication usage, laboratory results, and diagnoses. This rich information becomes even more valuable when organized temporally. To harness this potential, we created temporal trajectories from each patient's data, starting from their initial MCI diagnosis and extending until their first AD diagnosis. These generated temporal trajectories are described in the input data module of Figure 2A.

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.

To provide a mathematical framework for understanding the composition of our dataset, we describe it as a three‐dimensional tensor 𝜒, with dimensions 𝑛 × 𝑇 × 𝑑, where 𝑛 represents the total number of patients included in the dataset, each of whom is represented by an independent trajectory; 𝑇 signifies the number of time points recorded within the specified time window, reflecting the number of records commonly shared across most patient trajectories; and 𝑑 denotes the number of features recorded at each time point. Specifically, for any given patient 𝑖, where 𝑖 ∈ {1, 2, …, 𝑛}, their data are represented by the matrix 𝑋_𝑖, where 𝑋_𝑖 is a matrix of dimensions 𝑇 × 𝑑. Each row, 𝑥^𝑡 _𝑖, where 𝑡 ∈ {1, 2, …, 𝑇}, within this matrix corresponds to the 𝑑‐dimensional feature vector at that time point. Hence, the entire dataset 𝜒 can be expressed as Eq. (1):

$$\chi ={\left[\begin{array}{cccc}{X}_1,& X_2,& \cdots ,& X_n\end{array}\right]}^T.$$ (1)

In this framework, each 𝑋_𝑖 not only captures the patient's status at specific points in time but also encapsulates the evolution of the patient's state over time.

We applied the same data processing method across all sites. For continuous data, we preserved the original values after filtering out outliers and filled missing data with the average value of the feature. For discrete data, following the approach of Xu et al., ⁵¹ we used one‐hot encoding to encode discrete variables within our dataset. Given the diversity and complexity of discrete data features such as diagnosis codes and drug codes, encoding them could be challenging. To streamline the process and reduce complexity, we categorized these features: Diagnostic codes were mapped to Phecodes, allowing us to group complex diagnoses into several broad categories. This method was based on the system described in Bastarache. ⁵² Drug codes were matched to level 3 of the Anatomical Therapeutic Chemical (ATC) categorization system, including National Drug Codes (NDCs) and RxNorm. ⁵³ Using this categorization, drug data might be standardized according to chemical, pharmacological, and therapeutic qualities. After categorization, these features could be concatenated in a binary way by one‐hot encoding.

Inspired by Mao et al., ⁵⁴ we filtered the processed data into four cohort groups based on the time it took for their MCI to progress to AD. These groups were defined as progressing within 6 months, within 12 months, within 24 months, and without a time limit. Each group had specific requirements: subjects in the 6 month group needed at least 6 months of records after their initial MCI diagnosis, those in the 12 month group required at least 12 months of records, subjects in the 24 month group needed at least 24 months, and those in the “no time limit” group also needed a minimum of 24 months of records. The group information is shown in Table 2.

TABLE 2.

Total and case patients (i.e., converted to Alzheimer's disease) included across various follow‐up periods (i.e., 6 months, 12 months, 24 months, and no restrictions) at different sites.

	Total (cases)	North Florida (cases)	Central Florida (cases)	South Florida (cases)	Alabama (cases)	Georgia (cases)	Other (cases)
In 6 months
# Patients	18,060 (454)	2555 (71)	1994 (48)	1380 (21)	4314 (119)	179 (4)	7638 (191)
# Average convert days *	– (89.7)	– (92.3)	– (81.2)	– (105.4)	– (91.8)	– (129)	– (87.1)
# Average encounters	209.8 (194.2)	181.5 (152.7)	55.6 (64.0)	234.7 (202.5)	90.4 (87.9)	87.8 (106.2)	325.4 (310.0)
# Average encounter days	2979.7 (2865.6)	2930.8 (2836.6)	2584.2 (2521.4)	2930.4 (2581.8)	3030.2 (2983.6)	763.5 (762.0)	3131.8 (2964.6)
In 12 months
# Patients	15,468 (719)	2194 (100)	1621 (76)	1224 (40)	3729 (198)	37 (1)	6663 (304)
# Average convert days	– (175.3)	– (164.3)	– (169.4)	– (198.1)	– (185.5)	– (134)	– (171)
# Average encounters	223.9 (201.0)	190.8 (168.6)	56.4 (59.3)	245.5 (251.2)	94.1 (84.7)	94.8 (149.0)	345.0 (316.4)
# Average encounter days	3093.5 (2954.6)	3036.3 (3011.8)	2719.7 (2573.4)	3021.6 (2834.4)	3124.7 (3011.9)	811.3 (819.0)	3211.6 (3016.6)
In 24 months
# Patients	11,671 (911)	1703 (146)	1148 (92)	948 (45)	2759 (235)	0	5113 (393)
# Average convert days	– (340.2)	– (354.5)	– (302.2)	– (374.4)	– (348.7)	–	– (334.7)
# Average encounters	245.9 (225.3)	203.8 (184.8)	58.8 (62.3)	256.1 (211.2)	100.3 (81.8)	–	378.7 (366.0)
# Average encounter days	3241.8 (3081.3)	3210.0 (3093.8)	2872.9 (2760.2)	3187.4 (3083.9)	3276.2 (3054.0)	–	3326.7 (3167.9)
No restrictions
# Patients	12,648 (1888)	1841 (284)	1229 (173)	988 (85)	3005 (481)	0	5585 (865)
# Average convert days	– (876.7)	– (791.6)	– (772.7)	– (832.5)	– (893.6)	–	– (920.4)
# Average encounters	245.9 (235.5)	206.5 (211.0)	57.9 (53.9)	253.4 (201.0)	99.1 (84.2)	–	377.8 (367.5)
# Average encounter days	3249.0 (3212.6)	3219.7 (3213.1)	2976.5 (2838.7)	3197.6 (3251.7)	3281.5 (3200.9)	–	3332.2 (3289.8)

^*Average convert days: the average number of days that patients transition from mild cognitive impairment to Alzheimer's disease; “–”: N/A.

2.2. Models and experiment setting

In this subsection, we explain the workings of the LSTM and FL models, followed by a description of the experimental settings.

2.2.1. LSTM model for predicting MCI to AD conversion

To more accurately analyze these temporal trajectories, we have adopted a deep learning neural network based on LSTM. As shown in Figure 2A, the model is composed of two parts: a sequence of LSTM encoders and a predictor powered by a multilayer perceptron (MLP). As a classic time series algorithm, the sequence of LSTM encoders can gradually process patients’ EHRs through “memory cells” and store key information for a long time. This allows it to effectively capture dynamic changes in health status and predict disease progression from EHR data. ⁵⁵ After the data are extracted by a series of LSTM encoders, the extracted information is fed into the predictor to forecast whether the patient will eventually have AD. By minimizing the discrepancy between the ground truth and the predicted labels, the model undergoes continuous refinement until it converges.

2.2.2. Federated model aggregation

For FL systems, as shown in Figure 2B, we begin by generating a model with initial weights in the federation server, which is then distributed to each participating site machine. Each site then trains the LSTM model using its own private data. Upon completing 10 epochs, each site uploads its model weights back to the server. The server uses the federated averaging algorithm (FedAvg) ⁵⁶ to synthesize all the uploaded weights and redistribute the aggregated weights back to each site machine, replacing the original weights. This cycle constitutes one training round, and multiple rounds are run until the model converges.

2.2.3. Experiments setting

In this study, we established four experimental set‐ups across four distinct cohort groups. For the data, except for Georgia, we split the data from all sites according to the following ratios: training:validation:testing = 70:15:15. Due to the limited amount of data, all data from the Georgia site were used exclusively for testing. All experimental models used the best‐performing model from the validation process for subsequent testing. The prediction performance was measured using AUC. We implemented all models using Tensorflow 2.10.1, Tensorflow Federated Learning 0.40.0, and CUDA 11.6 by using a single NVIDIA A100 GPU with 40 GB RAM and 3 CPU cores.

1. Local model (baseline). This set of experiments used a single‐layer LSTM model with 64 hidden nodes, and the predictor consisted of one fully connected layer with sigmoid activation function. Each site used its unique private data for training and validation. We set the training duration for each site at 300 epochs to ensure comprehensive learning and model convergence. The training used binary cross‐entropy loss and an Adam optimizer with a learning rate of 0.001. After completion of training, each model was separately evaluated using its respective test set.

2. Pooling model. The pooling model was similar to the local model but combined the training and validation sets from all six sites. After training for 300 epochs, the model was independently tested on each of the six different test sets.

3. Global model. FL operated with both a global and a personalized model. As mentioned in section 2.2.2, one round involved downloading the weights, conducting local training for 10 epochs, and then uploading the weights to the server where they were aggregated using the FedAvg algorithm. After 200 rounds of training within the FL framework, the global model was obtained from the central server and tested on the six‐site test sets.

4. Personalized model. The personalized model was based on the global model. This set of experiments built on the global model by reintegrating the parameters into the LSTM models at each of the six sites. These models were then fine‐tuned through an additional 50 epochs of training and validation before being tested.

2.3. Evaluating feature importance across sites using SHAP and PFI analyses

We chose the two most populated sites, Alabama and Other Linked Data Regions, to assess feature importance across two distinct cohort groups: within 24 months and without restrictions. This analysis aimed to verify the models’ ability to handle diverse data characteristics by calculating the SHAP values. SHAP values were calculated based on Shapley values from game theory and represent the average contribution of a feature across all possible combinations of features. Essentially, the SHAP value was determined by comparing the models’ predictions with and without a particular feature. Because SHAP is not suitable for LSTM models, we used information from a single time point, the first MCI diagnose date, for each patient and applied XGBoost to calculate the SHAP values for the features in this dataset.

To further assess whether these features remain significant in the context of time‐series data, we applied PFI. Specifically, we evaluated feature relevance by comparing AUC values before and after permuting the data for individual features in both global and personalized models. Furthermore, we identified temporally important features and arranged them chronologically to present the trend of these features over time.

3. RESULTS

We identified a total of 44,899 patients from six sites, with a statistical summary presented in Table 1. The candidate cohort had an average age of 71.5 years, with 52.8% being women. Among the cohort, 80.4% were non‐Hispanic, and 76% were White. Population characteristics was generally similar across locations; however, two sites, South Florida and Georgia, had a higher percentage of Hispanic individuals (40%), compared to North Florida (5%). Racially, Georgia had a notably lower percentage of Whites (55%) than the other sites.

Among the study population, 14.2% (6391 patients) progressed from MCI to AD, while 85.8% (38,508 patients) remained in the MCI state throughout the study. In the South Florida site, only 9.1% MCI patients transitioned to AD, while other sites had a conversion rate of ≈ 14%, consistent with previous studies. ⁵ , ⁶ On average, patients had ≈ 140 encounters over ≈ 2100 days with notable variation in encounter frequencies and durations across different sites. We also documented the top 10 prevalent diseases and medications in Table 1. Predominantly, patients suffered from primary hypertension, hyperlipidemia, and neurological disorders, though the prevalence of these conditions was 10% to 30% lower in South Florida and Georgia compared to the general population, a trend also reflected in medication usage.

The patient counts for each follow‐up cohort are presented in Table 2. The “other linked sites” group had the largest population. Georgia, which only contains data from Emory University, has the smallest population, with no patients in the 24 month and “no time limit” groups. Additionally, the data indicated that as the follow‐up period increased, both the average number of encounters and the average days per encounter increased.

Table 3 presents the experimental results for various models predicting the progression from MCI to AD by site. As predicted, the personalized and pooled models performed best, followed by the global model, with the local model performing the poorest. The personalized and pooled models achieved the highest AUC scores (≈ 0.672 and ≈ 0.668, respectively), outperforming the global model (≈ 0.648) and the local model (≈ 0.619). Additionally, the uniformity in the findings across different cohort groups for each of the four datasets underscores the robust generalizability of the model. The consistency in results across different cohort groups for the four datasets underscores the robustness and generalizability of the models, highlighting their adaptability to different data characteristics.

TABLE 3.

AUC performance of different models for MCI to AD prediction by sites. Numbers in bold font represent the best performance, and numbers in italic font represent the second‐best performance. All available data from the Georgia site are used for testing due to data limitations.

	Model	North Florida	Central Florida	South Florida	Alabama	Georgia	Other
In 6 months	Local	0.729	0.579	0.533	0.690	–	0.655
	Pooled	0.693	0.616	0.637	0.721	0.784	0.712
	Global	0.785	0.622	0.57	0.715	0.531	0.662
	Personalized	0.794	0.651	0.606	0.712	–	0.679
In 12 months	Local	0.554	0.447	0.616	0.611	–	0.625
	Pooled	0.739	0.439	0.717	0.714	0.778	0.629
	Global	0.598	0.438	0.617	0.637	0.611	0.637
	Personalized	0.643	0.525	0.659	0.664	–	0.635
In 24 months	Local	0.580	0.532	0.820	0.692	–	0.642
	Pooled	0.593	0.638	0.811	0.723	–	0.657
	Global	0.609	0.637	0.837	0.755	–	0.652
	Personalized	0.572	0.621	0.836	0.763	–	0.657
No restrictions	Local	0.614	0.566	0.615	0.661	–	0.623
	Pooled	0.707	0.607	0.765	0.668	–	0.653
	Global	0.655	0.636	0.726	0.666	–	0.658
	Personalized	0.682	0.647	0.744	0.649	–	0.638

Abbreviations: AD, Alzheimer's disease; AUC, area under the receiver operating characteristic; MCI, mild cognitive impairment.

Figure 3 visualizes the importance of features across the two most populated sites and cohort groups using SHAP values. Each subplot shows the top 15 variables with the highest average SHAP values on the left, while the right‐side plot displays SHAP values for each feature for each sample, illustrating whether the feature is positively or negatively correlated with model predictions. Features such as body mass index (BMI), vitamin B12, systolic and diastolic blood pressure (BP), cholesterol, and high‐density lipoprotein (HDL) appeared consistently across the four data sets. These features were validated by PFI, as shown in Figure 4. On average, disrupting these features reduced the AUC by 1%, with the largest decrease observed in vitamin B12, for which the AUC dropped nearly 10% in both global and personalized models. Figure 5 presents the temporal trends of these six key features, with mean and standard deviation (SD) calculated every 3 months. The red line represents the case group (patients who transitioned from MCI to AD), while the blue line corresponds to the control group (patients who remained in the MCI stage). Notable differences in BMI, systolic and diastolic BP, and HDL levels were observed between the two groups, highlighting their potential role in the progression from MCI to AD. In contrast, vitamin B12 and cholesterol levels showed no substantial variations, particularly in Other Linked Data Regions.

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.

4. DISCUSSION

We aimed to demonstrate the feasibility of FL in predicting the transition from MCI to AD using EHRs and LSTM models. This approach not only protects data privacy but also uses decentralized data sources to improve model generalization and leverages the properties of time series models to enhance the ability to capture data. By reducing the requirement to send protected health information, we were able to avoid possible privacy and data governance restrictions in multicenter research by adopting a streamlined two‐way cooperation approach. Mitigating these difficulties by the implementation of FL helped speed the process of ethical and regulatory evaluation while maintaining patient privacy. ⁵⁷ Given that EHRs include patient treatment data at multiple time points, including clinical records, laboratory testing, and medication usage, the local model of LSTM allows for the concatenation of data from these time points, resulting in additional information.

In this study, our FL model demonstrated outstanding efficacy across the four different follow‐up periods, with overall performance comparable to the pooled model and better than any of the local models. This indicates that the FL model has been effective in predicting AD; at the same time, the FL model is able to profit from the data collected by each local model, all without the need for data merging, which considerably decreases the training cost.

In addition, our demographic analyses uncovered regional variations that are crucial for interpreting the study results. For instance, the proportion of non‐Hispanic individuals in North Florida reaches as high as 95%, in stark contrast to ≈ 60% in South Florida and Georgia. Moreover, the rate of MCI conversion to AD varied regionally, with rates as low as 9.1% in South Florida and ≈ 14% in other areas.

Six features (i.e., BMI, vitamin B12, systolic/diastolic BP, cholesterol, and HDL) in Figure 3 and Figure 4 were identified as having significant effects on the model across the four settings. Of these, BMI, diastolic BP, vitamin B12, and cholesterol were negatively associated with the transition from MCI to AD, whereas HDL and systolic BP were positively associated. These correlations align with existing research: Low BMI may reflect malnutrition or metabolic problems in the body, which are potential risk factors for AD. ⁵⁸ Similarly, low vitamin B12 levels may lead to reduced neurological function, ⁵⁹ and low diastolic BP may be associated with reduced blood flow to the brain. ⁶⁰ Abnormal cholesterol levels may affect the structure and function of nerve cells. ⁶¹ Higher HDL is often seen as a marker of cardiovascular health, but in the context of AD, it may reflect a complex imbalance in metabolic regulation associated with the disease. ⁶² Higher systolic BP may indicate the presence of ongoing cardiovascular stress or other factors that may accelerate cognitive deterioration by affecting brain structure and function. ⁶⁰ In addition, different variables may interact in complex ways. For instance, low BMI is often accompanied by weight loss, which has been strongly associated with vitamin B12 deficiency and neurodegenerative changes. ⁶³ Therefore, rather than simply reflecting malnutrition, low BMI may also indicate metabolic abnormalities that accelerate neurodegeneration. ⁶³ Likewise, the interplay between high systolic BP and elevated HDL levels may significantly impact neurodegenerative processes. Research suggests that hypertension and hypercholesterolemia not only increase the risk of dementia by inducing atherosclerosis and impairing blood flow but may also directly contribute to neurodegeneration in AD. ⁶⁴ In Figure 5, we can observe the temporal trends of these features. Because we used mean values to impute missing data and the SD is large, the trends between case and control groups might appear similar, especially for features with a high rate of missing values. This could explain the lack of significant differences in vitamin B12 and cholesterol levels in Other Linked Data Regions, as mentioned in the Results section. It may also contribute to the relatively modest AUC values observed in our models.

There are several limitations to this study. The first is its reliance on EHR data, which inherently comes with certain constraints. EHRs are primarily collected for clinical purposes rather than research, often leading to incomplete or inconsistent data. Variability in how health‐care providers document information, along with missing data or miscoded diagnoses, can introduce biases. Additionally, temporal gaps in EHR data may affect the accuracy of recorded disease progression or treatment outcomes, as patient follow‐up schedules vary. These factors could skew research results, reducing robustness and potentially impacting generalizability. In terms of data, we used diagnostic codes to identify patients with MCI and AD. However, MCI is an intermediate stage between normal cognitive function and AD, and its progression varies between individuals. Before 2010, AD diagnoses were typically not given unless clear dementia symptoms were present. ⁶⁵ With the advent of in vivo biomarkers, physicians have begun to separate disease severity (or syndrome) from disease etiology. ⁶⁵ For example, MCI progressing to dementia should now be specified as dementia due to AD. Although clinical coding is gradually adapting to this conceptual shift, many systems still retain pre‐2010 terminology, equating AD with dementia, as seen in ICD codes and the Diagnostic and Statistical Manual of Mental Disorders (DSM). ⁶⁶ This may lead to misclassification of patients as having either MCI or AD. Additionally, certain features used in our analysis may introduce bias. For instance, BMI was treated as a continuous variable, which may have influenced the results. A large body of literature suggests a U‐shaped association between BMI and dementia risk, where both abnormally low and high BMI are linked to increased risk. ⁶⁷ Treating BMI as a continuous variable might have obscured this non‐linear relationship and impacted our findings. Furthermore, our study was limited to data from OneFlorida+, which primarily includes data from Florida, Alabama, and Georgia (with a relatively small contribution from Georgia). The geographic restriction may limit the generalizability of our findings to populations with different demographics. Moreover, our study was limited to structured data, potentially overlooking nuanced insights from physicians that could be captured in unstructured modality like clinical notes. Last, our modeling approach used a single‐layer LSTM model as the local model. While effective in capturing sequential patterns, this simple architecture may not have been sufficient to fully extract complex relationships within the data.

5. CONCLUSION

This study demonstrates the potential efficacy of FL using real‐world EHR data from patients with MCI and AD, collected from six regional sites across four different settings. FL technology provides a promising approach to maintaining patient privacy while enabling data integration across multiple organizations. Additionally, it allows for an assessment of how data heterogeneity across different regions influences model performance. The findings suggest that FL could be a valuable tool for clinical analysis of AD, with potential applications on a global scale. Further research is needed to deepen our understanding of AD mechanisms and to explore early diagnostic and preventive strategies.

AUTHOR CONTRIBUTIONS

Jinqian Pan and Jie Xu were responsible for the overall design, development, and evaluation of this study. Jinqian Pan, Zhengkang Fan, and Jie Xu did the initial drafts of the manuscript. Zhengkang Fan was responsible for running all experiments and creating SHAP plots. All authors reviewed and revised the manuscript critically for scientific content.

CONFLICT OF INTEREST STATEMENT

The authors declare no competing interests. Author disclosures are available in the supporting information.

CONSENT STATEMENT

The study has been approved and the requirement to obtain any informed consent has been waived by the University of Florida Institutional Review Board (protocol no. IRB202202820). The research does not involve greater than minimal risk for participation. Analyses only involve the secondary analysis of data that are either limited data sets or de‐identified. Our research team has no direct contact with human subjects. All methods were carried out in accordance with relevant guidelines and regulations.

Supporting information

Supporting information

Pan J, Fan Z, Smith GE, Guo Y, Bian J, Xu J. Federated learning with multi‐cohort real‐world data for predicting the progression from mild cognitive impairment to Alzheimer's disease. Alzheimer's Dement. 2025;21:e70128. 10.1002/alz.70128