Identifying Progression Subphenotypes of Alzheimer’s Disease from Large-Scale Electronic Health Records with Machine Learning

Abstract

Objective

Identification of clinically meaningful subphenotypes of disease progression can enhance the understanding of disease heterogeneity and underlying pathophysiology. In this study, we propose a machine learning framework to identify subphenotypes of Alzheimer’s disease progression based on longitudinal real-world patient records.

Methods

The framework, dynaPhenoM, extracts coherent clinical topics across patient visits and employs a time-aware latent class analysis to characterize subphenotypes. We validated dynaPhenoM using three patient databases with a total of 3,952 AD patients across the United States, demonstrating its effectiveness in revealing mild cognitive impairment (MCI) progression to AD.

Results

Our study identified five subphenotypes associated with distinct organ systems for disease progression from MCI to AD, including common subtypes across cohorts—respiratory, musculoskeletal, cardiovascular, and endocrine/metabolic—as well as a cohort-specific digestive subtype.

Conclusion

Our study unravels the complexity and heterogeneity of the progression from MCI to AD. These findings highlight disease progression heterogeneity and can inform both diagnostic and therapeutic strategies, thereby advancing precision medicine for Alzheimer’s disease.

1. Introduction

Alzheimer's disease (AD) is the most prevalent neurodegenerative disorder worldwide[1], with its prevalence expected to double in the next 20 years [1,2]. The disease mechanism of AD is highly complex, and effective treatments are yet to be found[3]. The complex and heterogeneous nature of human diseases, such as AD, often results in patients exhibiting diverse clinical manifestations[4]. Identifying clinically meaningful subphenotypes, or subgroups of patients with coherent clinical characteristics, is crucial for enhancing our understanding of underlying disease mechanisms and informing precision medicine[5,6]. With the increasing adoption of health information systems such as electronic health records (EHR), comprehensive patient information, including demographics, diagnoses, medications, and lab tests, have become readily available[7].

To effectively explore clinical information in patient records and identify comprehensive disease subphenotypes, several challenges must be addressed: (i) information heterogeneity: patient data encompass various types of information; (ii) irregular timing of visits: the time intervals between successive patient visits are typically irregular; (iii) missing values: substantial information may be missing in patient records, which does not necessarily indicate the absence of the disease; (iv) high-dimensionality and sparsity: clinical events are represented as systematic codes with large vocabularies[12], with each patient visit containing only a few codes[13]; and (v) interpretability: analysis results must be interpretable and easily understood by clinicians.

While existing studies have developed data-driven approaches to identify and predict disease subphenotypes[8-11], they often focus on selected clinical events without considering their temporal evolution. For instance, traditional machine learning models such as random forests and XGBoost have been used to classify Alzheimer’s subtypes[10] but did not consider disease trajectories over time to model AD progression. In addition, methods such as LSTM-based models and Hidden Markov Models, have been developed to model temporal dynamics[11] using patients’ longitudinal records but usually suffer from limitations in interpretability and scalability. To address the issues, in this study, we developed a machine learning framework, termed dynaPhenoM, which capture temporal patterns of AD progression using longitudinal patient records to identify subphenotypes while ensuring their clinical interpretability. The dynaPhenoM comprises two main modules: the dynamic multimodal topic model (DMTM) for deriving interpretable compressed representations of multimodal clinical events, and the time-aware latent class analysis (TLCA) for subphenotype identification that accommodates irregular visit times. DMTM builds on latent topic modeling (LTM)[14], commonly used in text-mining tasks, treating clinical events as words and each visit as a document. Unlike existing methods that focus on single visits, DMTM learns representations from longitudinal patient information[12,15-17]. Although latent class analysis (LCA)[17] is a widely used subphenotyping method in clinical studies[18-20], traditional LCA does not account for irregular time intervals between visits[21], motivating the invention of TLCA. We further developed and applied machine learning models to predict subphenotype assignment based on clinical features prior to MCI onset.

To demonstrate effectiveness of the dynaPhenoM, we applied it to identify progression subphenotypes for patients progressing from mild cognitive impairment (MCI) to AD. MCI is an intermediate stage between normal aging and dementia, including AD[22]. Understanding the clinical heterogeneity of patients progressing from MCI to AD and identifying corresponding progression subphenotypes can potentially reveal different underlying disease pathophysiologies, improve diagnostics (not all people with MCI will progress to AD), and inform effective treatments. We leveraged three large-scale, real-world electronic health record (EHR) databases: the INSIGHT network, which consists of 2,579 AD patients with 10,562 clinical codes; the OneFlorida+ network, which consists of 950 AD patients with 9,364 clinical codes; and the EHR repository of University of California San Francisco (UCSF), which consists of 678 AD patients with 16,740 clinical codes. Both the INSIGHT and OneFlorida+ networks are part of the National Patient-Centered Clinical Research Network (PCORnet). Utilizing the INSIGHT dataset as the development cohort, we identified five clinically meaningful subphenotypes associated with different organ systems—respiratory, musculoskeletal, cardiovascular, endocrine/metabolic, and digestive subtype—highlighting the heterogeneous progression procedures from MCI to AD. To assess robustness and generalizability of our approaches and findings, we used the OneFlorida+ and USCF data as independent validation cohorts to replicate our analyses. We found that, out of the five subphenotypes identified in INSIGHT data, four are consistent across the validation cohorts, despite their distinct patient populations compared to the development cohort. Additionally, we conducted extensive statistical analysis and integrated relevant literature and clinical insights to strengthen the biological and clinical interpretability of our conclusions. Furthermore, applying subphenotype prediction modeling based on machine learning enables the forecasting of the progression trajectory from MCI to AD. Our findings contributed to the development of targeted diagnostic and treatment strategies, offering new opportunities for precision medicine in AD.

2. Methods

This study was approved by the Institutional Review Board (IRB) of Weill Cornell Medicine with protocol number 21-07023759. The use of OneFlorida+ data for this study is approved under the University of Florida IRB number IRB202001888. This study complies with all relevant ethical regulations and is approved by the Institutional Review Board of UCSF (IRB 20-32422).

2.1. Study cohorts

We conducted the study on 3 cohorts. We identified patients with AD and MCI based on ICD-9 and ICD-10 codes (Table S1). Initially, we included patients diagnosed with both MCI and AD. Subsequently, we excluded patients whose first AD onset time or dementia onset time preceded the MCI onset time. We aggregated all records within every three months as an “observation window”. We excluded patients who had only one observation window or were younger than 50 years old at MCI onset. As a result, we included 950, 678, and 2579 patients from OneFlorida+, UCSF, and INSIGHT who were diagnosed with AD, respectively (Figure S1a). Within the OneFlorida+ cohort, the median age was 77, and the progression time from MCI onset to AD onset was up to 8 years. The proportions of white, black, and other race patients were 50.53%, 13.15%, and 35.79%, respectively (Table 1, Table S2). Within the UCSF cohort, the median age was 77, and the progression time from MCI onset to AD onset was up to 16 years. The proportions of white, black, and other race patients were 66.81%, 6.34%, and 10.47%, respectively (Table 1). Within the INSIGHT cohort, the median age was 78, and the progression time from MCI onset to AD onset was up to 19 years. The proportions of white and black patients were 56.8% and 9.53%, respectively (Table 1). In this study, we used the INSIGHT as the model development cohort and used OneFlorida+ and UCSF as validation cohorts.

Table 1:

Characteristics Statistics of 3 cohorts. The Phecode of comorbidities is in Table S2.

	UCSF	OneFlorida+	INSIGHT
No. of patient	678	950	2579
MCI onset Age, median (IQR)	77 ([71, 84])	77 ([69,83])	78 ([72.0,83.0])
Sex female, N (%)	388 (57.22)	615 (64.73)	1577 (61.148)
Race, N (%)
Caucasian	453 (66.814)	480 (50.526)	1466 (56.844)
African American	43 (6.342)	125 (13.158)	246 (9.539)
asian	111 (16.372)	5 (0.526)	42 (1.629)
Other	71 (10.472)	340 (35.789)	825 (31.989)
Progression time, N (%)
1~2 year	255 (37.611)	255 (26.842)	742 (28.771)
2~3 year	147 (21.681)	195 (20.526)	593 (22.993)
3~4 year	102 (15.044)	185 (19.474)	401 (15.549)
4~5 year	59 (8.702)	161 (16.947)	275 (10.663)
>5 year	115 (16.962)	154 (16.211)	568 (22.024)
Comorbidity, N (%)
Hypertension	234 (34.513)	417 (43.895)	895 (34.703)
Hyperlipidemia	164 (24.189)	249 (26.211)	525 (20.357)
Diabetes	86 (12.684)	231 (24.316)	333 (12.912)
Memory loss	149 (21.976)	121 (12.737)	670 (25.979)
Heart disease	78 (11.504)	73 (7.684)	192 (7.445)
Sleep disorders	36 (5.310)	18 (1.895)	91 (3.528)
Anxiety	39 (5.752)	84 (8.842)	178 (6.902)
Gastroesophageal reflux disease	39 (5.752)	102 (10.737)	176 (6.824)
Cerebrovascular disease	12 (1.770)	55 (5.789)	60 (2.326)
Chronic airway obstruction	20 (2.950)	87 (9.158)	70 (2.714)
Chronic renal failure	44 (6.490)	35 (3.684)	90 (3.490)
Urinary tract infection	51 (7.522)	109 (11.474)	92 (3.567)
Glaucoma and Cataract	90 (13.274)	124 (13.053)	196 (7.600)
Medicine, N (%)
Antithrombotic agents	100 (14.749)	102 (10.737)	505 (19.581)
Gastrointestinal agents	72 (10.619)	24 (2.526)	47 (1.822)
Opioids	140 (20.649)	65 (6.842)	209 (8.104)
Antidepressants	139 (20.501)	133 (14.000)	457 (17.720)
Antiinfectives	159 (23.451)	113 (11.895)	521 (20.202)
Corticosteroids	178 (26.254)	125 (13.158)	453 (17.565)
Beta blocking agents	111 (16.372)	104 (10.947)	387 (15.006)
Anti-dementia drugs	68 (10.029)	64 (6.737)	218 (8.453)
Hypnotics and sedatives	136 (20.059)	60 (6.316)	174 (6.747)
Urological	96 (14.159)	49 (5.158)	214 (8.298)
Insulins and analogues	21 (3.097)	42 (4.421)	113 (4.382)

For the INSIGHT and the UCSF datasets, the patient diagnosis codes were recorded as ICD-9 and ICD-10 codes, medications were encoded by National Drug Code (NDC) codes, and procedures were encoded with Current Procedural Terminology (CPT) codes. For the OneFlorida+ dataset, the patient diagnosis codes were recorded as ICD-9 and ICD-10 codes, medications were encoded by NDC and RxCUI codes, and procedures were encoded with CPT codes. We then mapped diagnosis codes to Phecode[23] and normalized NDC codes to RxCUI.

2.2. DMTM model

We represented the diagnosis, drug, and procedure events as three binary feature vectors. Due to the large vocabulary size (total unique codes) of three modalities, the original feature vectors are high-dimensional and sparse[13], which makes efficient clustering difficult. Therefore, we proposed a novel probabilistic model, dynamic multimodal topic model (DMTM), to extract low-dimensional continuous features from original high-dimensional binary vectors. The new features extracted from DMTM are not only beneficial for the following derivation of subphenotypes, but also explainable for the exploration of subphenotypes.

As shown in Figure 1, DMTM models longitudinal multimodal clinical events from longitudinal patient events as a latent generative process, from the first observation window to the last, whose specific notations are provided in Table S3. After collecting all patient records, the $m$-th modality of $n$-th patient at $t$-th observation window can be represented as a binary vector $x_{n,t}^{(m)}\in {\{0,1\}}^{V_m}(m=1,\cdots ,M;t=1,\cdots ,T_n;M=3$ in our current case including diagnosis, medication, and procedure events), where $V_m$ represents the total unique clinical events (vocabulary size) in $m$-th modality, and $T_n$ is the total number of observation windows for the $n$-th patient. Suppose that there are $K$ latent clinical topics and each contains $M$ different types of topics corresponding to different modalities denoted as ${\Phi }^{(m)}\in R_{+}^{V_m\times K}$, in which the $k$-th column, ${\Phi }_k^{(m)}\in R_{+}^{V_m}$, represents $k$-th topic, a distribution over all events (unique codes) in $m$-th modality. DMTM assumes that $x_{n,t}^{(m)}$ is composed of $K$ topics with ${\theta }_{n,t}\in R_{+}^K$ being the topic weight vector (mixture composition) shared by all modalities. Therefore, $k$-th topic in different modalities $({\varphi }_k^{(1)},\cdots ,{\varphi }_k^{(M)})$ are highly correlated, forming as $k$-th clinical topics shown in Figure 2. To model the transition pattern of topic weights between two successive observation windows, DMTM introduces a transition matrix $\Pi \in R_{+}^{K\times K}$, where each element, ${\pi }_{ij}$, represents the probability of transition from $i$-th topic to the $j$-th topic. Formally, the generative process of DMTM can be written as:

Figure 1:

Data-driven identification of Alzheimer’s disease progression subphenotypes. (a) Electronic Health Record (EHR) database. We utilized 3 EHR databases across the country: INSIGHT (development cohort), OneFlorida+ (validation cohort), and UCSF (validation cohort). (b) Data preprocessing pipeline. We extracted data on diagnoses, medications, and procedures related to the progression from mild cognitive impairment (MCI) to Alzheimer's disease (AD). (c) Three outputs of the Dynamic multi-modal topic modeling. (d) Interpretable organ systems related topic patterns based on the clinical topic matrix. (e) Alzheimer’s disease subphenotyping through time-aware latent class analysis (T-CLA). (f) progression trajectories of organ systems-associated subphenotypes. (g) Early prediction of subtype assignment for new patients based on prediction models.

Figure 2:

Topic patterns and topic transition probabilities. (a-g) Circular plot of topic patterns. Within each circle, the green, purple, and red segments denote the top 10 diagnoses, medications, and procedures, respectively. The bar height is proportional to the topic weight, and the bar color signifies distinct organ system diseases. (h) Transition probabilities between key topics. Each node represents a topic, and the edges between nodes represent the transition probabilities. Acronyms: T2DM (Type 2 diabetes mellitus), NOS (not otherwise specified), CKD (chronic kidney disease), NEC (not elsewhere classified), CT (computed tomography), MRI (magnetic resonance imaging), ECG (electrocardiogram), vit B12 (vitamin B12), PTT (partial thromboplastin time), neuro (neurologic), ophtho (ophthalmologic), rheum (rheumatic).

• Topic weights: ${\theta }_{n,1}\sim Gamma(r,1),{\theta }_{n,t}\sim Gamma({\tau }_0\Pi {\theta }_{n,t-1},{\tau }_0),t=2,\cdots ,T_n$

• Latent clinical topics and transition matrix: ${\varphi }_k^{(m)}\sim Dirichlet({\eta }_0),{\pi }_k\sim Dirichlet(v_1{v}_k,\cdots ,\xi v_k,\cdots ,v_K{v}_k),k=1,\cdots ,K$

• Intermediate variables $v_k\sim Gamma\left(\frac{{\gamma }_0}{K},\beta \right),k=1,\cdots ,K;\xi ,\beta \sim Gamma({\mathit{\in }}_0,{\mathit{\in }}_0)$

• EHR clinical events which are represented as:

$$\begin{gathered} x_{n,t}^{(m)}=1\left(u_{n,t}^{(m)}\ge 1\right), \\ {u}_{n,t}^{(m)}\sim Poisson({\Phi }^{(m)}{\theta }_{n,t}),n=1,\cdots ,N;m=1,\cdots ,M;t=1,\cdots ,T_n, \end{gathered}$$

where, Gamma, Dihichlet, and Poisson denote the Gamma, Dirichlet, and Poisson distribution, respectively; $1(\cdot )$ is an indicator function representing that $x_{n,t}^{(m)}=1$ if $u_{n,t}^{(m)}\ge 1$, and $x_{n,t}^{(m)}=0$ if $u_{n,t}^{(m)}=0$. This function is called Bernoulli-Poisson link[24], whose mathematical motivation is that after transforming a binary-modeling problem (clinical event happens or does not happen in this observation window) into a count modeling one, one is readily equipped with a rich set of statistical tools developed for count data analysis using the Poisson and negative binomial distributions.

There are four positive hyperparameters to be set by users: ${\tau }_0,{\gamma }_0,{\eta }_0,{\in }_0$. In our setting, we set them as ${\tau }_0=1,{\gamma }_0=100,{\eta }_0=0.01,{\mathit{\in }}_0=0.1$. We developed the Gibbs sampling to estimate the posterior of all variables (in Supplementary Methods). Here, we only showed the posterior of topic weights ${\theta }_{n,t}$, to explain why DMTM can alleviate the problem of missing events in longitudinal patient records.

2.3. TLCA model

Most existing works in clinical research about deriving longitudinal subphenotypes were implemented using latent class analysis (group based trajectory modeling)[25] or dynamic time warping[26], which often regarded visit times rather than calendar time as time stamps. However, such methods ignored the fact that the time interval between two visits may be irregular, varying from days to months, which is important for clinical study since it embeds the progressive speed of diseases. To this end, in this paper, we introduced time-aware latent class analysis (TLCA).

Specifically, the new features extracted by DMTM are topic weight vectors denoted as $\Theta ={\{{\theta }_{n,t}\in R_{+}^K\}}_{n=1,t=1}^{N,T_n}$ (note that here we use $K$ to represent the number of key topics), TLCA models the data likelihood of $\Theta$ by a mixture of Gaussian distribution as:

$$pp(\Theta )={\sum }_n{\sum }_c{\alpha }_c{\prod }_{t=1}^{T_n}{\prod }_{k=1}^K𝒩({\theta }_{n,t,k}\mid {\beta }_{c,k}{\tau }_{\mathit{n},t},{\sigma }_k),$$

where, ${\{{\alpha }_c\}}_{c=1}^C$ are the mixture coefficients with $C$ being the number of subphenotypes, and the mean ${\beta }_{c,k}{\tau }_{n,t}$ is defined as:

$${\beta }_{c,k}{\tau }_{n,t}=[{\beta }_{c,k,1},{\beta }_{c,k,2},{\beta }_{c,k,3},{\beta }_{c,k,4}]{\left[1,d_{n,t}-d_{n,1},{(d_{n,t}-d_{n,1})}^2,{(d_{n,t}-d_{n,1})}^3\right]}^T,$$

where, $d_{n,1}$ is the calendar time of starting point (such as MCI onset in our case) for $n$-th patient; $d_{n,t}$ is the calendar time of $t$-th observation window for $n$-th patient. In other words, $d_{n,t}-d_{n,1}$ models the calendar time interval from the starting point to every observation window, which embeds the natural time progression. This is the reason why we call our proposed new type of LCA as time-aware LCA. To learn the parameters ${\{{\beta }_{c,k}\in R^4\}}_{c=1,k=1}^{C,K}$ and ${\{{\sigma }_k\in R^1\}}_{k=1}^K$, and infer the subphenotype belonging for each patient, we use the Expectation–Maximization (EM) algorithm[27], which are detailed in Supplementary Methods.

2.4. Predictive modeling of subphenotypes

We next turned to build predictive models of subphenotypes. Our goal is to distinguish the identified subphenotypes at an early stage. To this end, we leveraged patient records including diagnosis codes, medications, and procedure codes collected at within a 6-month window before MCI onset as candidate predictors. Since we have more than 2 subphenotypes, we treated the task, prediction of patients’ subphenotype memberships, as a multi-classification problem. Therefore, we trained the model using a one-vs-rest strategy, i.e., the model contains a set of base classifiers, each of which learns to separate one specific subphenotype out from the others. Here we employed logistic regression as the base model. Within each cohort, we randomly split the data into 80% training set and 20% testing set. The prediction model was trained using the training set and evaluated using the testing set. We repeated above procedure 10 times.

To measure model performance, we used the metrices including accuracy and Micro Area under the Operating Characteristics Curve (Micro-AUC). Accuracy is defined as the proportion of patients whose subphenotype memberships were correctly predicted across the whole population. To calculate the Micro-AUC, we first obtained the true positive rate (TPR) and false positive rate (FPR) for each subphenotype-vs-rest base model. We then aggregated all TPRs and FPRs to generate a single Receiver Operating Characteristic (ROC) curve, from which the AUC was computed.

3. Results

3.1. Overall pipeline

The overall pipeline is illustrated in Figure 1. We first extracted patient records in terms of diagnoses, medications, and procedures within the period of progression from MCI to AD (Figure 1a, Figure S1a, Table S1). To overcome the challenge of sparse longitudinal medical events in EHR data, we divided the study period into smaller observation windows, each spanning three months, and aggregated patients’ records within these observation windows (Figure 1b). This approach, inspired by previous studies[28], was designed to balance two competing factors: minimizing noise associated with short time intervals while reducing the risk of missing important health condition trends that may emerge over longer periods. For every observation window, records from one modality (diagnoses, medications, or procedures) are represented as a binary vector, where 1 indicates the inclusion of a specific code and 0 indicates its absence. The length of this binary vector is equal to the total number of unique codes in the modality. In the INSIGHT dataset, we used 1,554 diagnosis codes (Phecode)[29], 6,384 medication codes (RxNorm)[29,30], and 2,624 procedure codes (CPT)[31]. In the OneFlorida+ and UCSF datasets, we used 1,513 and 1,421 diagnosis codes (Phecode), 4,152 and 13,532 medication codes (RxNorm), and 3,699 and 1,784 procedure codes (CPT), respectively. We identified patients with AD and MCI based on ICD-9 and ICD-10 codes (Table S1). We Initially included patients diagnosed with both MCI and AD. Subsequently, for clinical clarity, we excluded patients whose first AD onset time or dementia onset time preceded the MCI diagnosis onset time. We also excluded patients who had only one observation window or were younger than 50 years old. As a result, we retained 2,579 patients in the INSIGHT cohort, 950 patients in the OneFlorida+ cohort, and 678 patients in the UCSF cohort. The demographic information, including age, race, gender, and progression time from MCI to AD diagnosis, along with comorbidity percentage and medication usage at baseline (1 year prior to MCI onset), is summarized in Table 1.

Next, we trained a topic model on this multimodal data to learn a set of multimodal clinical topics from the patient records separately in each cohort (Figure 1c). Each topic can be viewed as a set of clinical events that are more likely to co-occur within an observation window (i.e. aggregated 3-months visits). For every observation window, we inferred the mixture memberships, also called topic weights, as the new representations of this observation window. A higher weight on a particular topic indicated that the corresponding observation window included more events from this topic. This topic weight-based representation transformed the visits representation from the original high-dimensional binary space to a low-dimensional continuous space, where each dimension was a topic composed of a set of frequently co-occurring clinical events. Thus, this representation was highly interpretable. By concatenating the learned representations for all observation windows of a specific patient according to the timeline, we obtained a multivariate temporal sequence for the patient with irregular intervals. DMTM learned new representations of observation windows through a dynamic multi-modal topic modeling process.

Finally, we clustered the topics to identify distinct subphenotypes (Figure 1d). Considering the irregular patient visits, TLCA derived the progression subphenotypes from the trajectories of these new representations through an LCA type of process. We then derived disease progression subphenotypes by grouping these temporal sequences with TLCA.

3.2. DMTM Extracted Interpretable and Meaningful Topics Relevant in AD Progression

We applied the DMTM model to the diagnosis, medication, and procedure data and identified distinct coherent topics in the INSIGHT cohort (Methods, Figure S1b). In topic modeling, a challenge is to determine the optimal topic number within the data. To address this, we calculated Topic Coherence Score, which measures the semantic coherence of the top clinical codes within the identified topics, with a higher value indicating better topic structure within the data. Additionally, we also considered clinical interpretation of the identified topics. Based on these two notions, we found 30 hidden topics within the data (Figure S2a). Subsequently, the coherent topics were identified based on the prominence of diagnosis categories (Figure 2, Figure S3). Among these, we identified 13 coherent topics, with each coherent topic characterized by a unique disease based on the Phecode categories[12].

Figure 2 shows the topic patterns learned from the INSIGHT cohort. Each circle within the plot encompasses three key components: diagnosis events, medicine usage, and medical procedures. The height of the vertical bars within each circle corresponds to the topic weight, indicating the relative importance of each component. Distinct colors are employed to differentiate various organ system diseases, allowing for a visual representation of the relative composition of different diseases. We observed that each topic was typically related to one organ system. For example,

• T10 (Figure 2c) is associated with sensory organ diseases, with all five of its top diagnoses related to the eyes (senile cataract, glaucoma, cataract, primary open angle glaucoma, disorders of vitreous body) based on the Phecode categories[12].

• Topic T14 (Figure 2d) is predominantly characterized by cardiovascular diseases, with 4 out of its top 5 diagnosis events corresponding to heart-related conditions (atrial fibrillation and flutter, coronary atherosclerosis, heart failure NOS, and nonrheumatic mitral valve disorders).

• T22 (Figure 2e) is associated with genitourinary diseases, with all five of its top diagnoses related to the genitourinary system (hyperplasia of prostate, frequency of urination and polyuria, retention of urine, other abnormality of urination, urinary incontinence).

• T23 (Figure 2f) is dominated by respiratory system diseases, especially lung cancer issues. It includes diagnosis events such as abnormal findings on examination of lungs, emphysema, chronic airway obstruction, shortness of breath, and cough.

• Topic T26 (Figure 2g) exhibits a predominant association with digestive diseases, as 4 out of its top 5 diagnoses are digestive diseases (abdominal pain, diarrhea, irritable bowel syndrome, diverticulosis).

We observed strong coherence across the three modalities for each topic. In genitourinary related Topic T22, 4 of its top 5 medicines are targeted at addressing genitourinary diseases (tamsulosin, finasteride, tamsulosin, dutasteride), and 2 out of its top 5 medical procedures (bladder capacity measurement and complex uroflowmetry) are also linked to genitourinary conditions. Tamsulosin, an alpha-1 blocker, was approved for the treatment of benign prostatic hyperplasia, which describes a proliferative process of both stromal and epithelial elements of the prostate[32]. Likewise, in cardiovascular related Topic T14, all five top medications target the cardiovascular system (apixaban, metoprolol, furosemide, lipitor) based on the Anatomical Therapeutic Chemical (ATC) 1st level classification. Apixaban is a novel oral anticoagulant for the prevention of stroke and blood clots in patients with nonvalvular atrial fibrillation[33]. Furthermore, 2 of its top 10 procedures (electrocardiogram, complete transthoracic echocardiography) evaluate the electrophysiology and structure of the heart. In the respiratory related Topic T23, 1 out of top 5 medications, tiotropium, targeted chronic obstructive pulmonary disease. Additionally, in T10, 4 out of top 5 procedures (ophthalmological exam, retinal imaging, extended visual field exam, extended ophthalmoscopy) are utilized to examine eye conditions.

We showed the transition probability between different topics, revealing a tendency for disease topics to remain consistent across consecutive patient visits. In Figure 2i, each entity represents a topic colored by its dominant disease, and the flow signifies the transition probability between topics with time. The majority of the flow returning to the original entity suggested a likelihood of topics remaining undisturbed. In addition, certain topics persisted with relatively higher values. For instance, there was a 13.7% transition probability from anemia-related diseases (T21) to digestive diseases (T26) (Figure 2i). Indeed, anemia is a common extraintestinal manifestation of inflammatory bowel disease, which is inflammation of the digestive system[34]. Therefore, DMTM learned interpretable and clinically meaningful topics patterns, revealed possible disease transition probability, and facilitated the subsequent derivations of progression subphenotypes.

3.3. TLCA Identifies Subphenotypes of AD Progression Based on Organ Systems

After identifying coherent AD progression topics, we characterized patients based on these topics and derived potential AD progression subphenotypes as clusters of patients with similar topic profiles (See Methods). To determine the optimal number of clusters, we evaluated different numbers of clusters and selected the optimal cluster number as 5 using NBclust[35] (Figure S2b). Consequently, five distinct subphenotypes were derived from the INSIGHT cohort. Table 2 summarizes their characteristics with respect to patient demographics, medical utilizations and pre-existing conditions in the baseline period (1 year before MCI onset). We characterized these subphenotypes in detail as follows (Figure 3, Figure S3).

Table 2:

Subphenotype statistics of the INSIGHT cohort. The p-value for sex, race, key comorbidities and medicines are obtained by χ2 test. The p-value for age and progression time are obtained by Kruskal-Wallis test.

	Subtype 1	Subtype 2	Subtype 3	Subtype 4	Subtype 5	p value
No. of patient	393	253	509	725	699
Age (MCI onset), Median (IQR)	80 ([75.0,85.0])	77 ([72.0,83.0])	78 ([73.0,83.0])	79 ([73.0,84.0])	75 ([69.0,81.0])	<0.0001
Sex female, N (%)	284 (72.26)	163 (64.42)	290 (56.97)	458 (63.17)	382 (54.64)	<0.0001
Race, N (%)						<0.0001
Caucasian	172 (43.76)	115 (45.45)	340 (66.79)	342 (47.17)	497 (71.1)
African American	24 (6.106)	36 (14.22)	39 (7.662)	117 (16.13)	30 (4.291)
asian	5 (1.272)	3 (1.185)	13 (2.554)	15 (2.068)	6 (0.858)
Other/Unknown	192 (48.85)	99 (39.13)	117 (22.99)	251 (34.62)	166 (23.75)
Progression time, Median (IQR)						<0.0001
1~2 year, N (%)	86 (21.88)	69 (27.27)	185 (36.34)	203 (28)	199 (28.46)	<0.0001
2~3 year, N (%	87 (22.13)	49 (19.36)	139 (27.3)	164 (22.62)	154 (22.03)	0.099
3~4 year, N (%)	61 (15.52)	31 (12.25)	69 (13.55)	116 (16)	124 (17.73)	0.178
4~5 year, N (%)	48 (12.21)	25 (9.881)	47 (9.233)	77 (10.62)	78 (11.15)	0.658
>5 year, N (%)	111 (28.24)	79 (31.22)	69 (13.55)	165 (22.75)	144 (20.6)	<0.0001
Comorbidity, N (%)
Hypertension	172 (43.77)	96 (37.94)	217 (42.63)	296 (40.83)	114 (16.31)	<0.0001
Hyperlipidemia	72 (18.32)	51 (20.16)	167 (32.81)	168 (23.17)	67 (9.59)	<0.0001
Diabetes	37 (9.41)	41 (16.21)	109 (21.41)	113 (15.59)	33 (4.72)	<0.0001
Memory loss	114 (29.01)	65 (25.69)	135 (26.52)	163 (22.48)	193 (27.61)	0.110
Heart disease	20 (5.09)	21 (8.30)	69 (13.56)	66 (9.10)	16 (2.29)	<0.0001
Sleep disorders	15 (3.82)	11 (4.35)	26 (5.11)	14 (1.93)	25 (3.58)	0.045
Anxiety	40 (10.18)	19 (7.51)	31 (6.09)	43 (5.93)	45 (6.44)	0.074
Gastroesophageal reflux disease	36 (9.16)	21 (8.30)	45 (8.84)	42 (5.79)	32 (4.58)	0.007
Cerebrovascular disease	6 (1.53)	4 (1.58)	16 (3.14)	23 (3.17)	11 (1.57)	0.123
Chronic airway obstruction	7 (1.78)	3 (1.19)	22 (4.32)	19 (2.62)	19 (2.72)	0.074
Chronic renal failure	12 (3.05)	7 (2.77)	22 (4.32)	44 (6.07)	5 (0.72)	<0.0001
Urinary tract infection	10 (2.54)	9 (3.56)	25 (4.91)	30 (4.14)	18 (2.58)	0.161
Glaucoma and Cataract	25 (6.36)	87 (34.39)	19 (3.73)	49 (6.76)	16 (2.29)	<0.0001
Medicine, N (%)
Antithrombotic agents	95 (24.17)	46 (18.18)	121 (23.77)	147 (20.28)	96 (13.73)	<0.0001
Gastrointestinal agents	4 (1.02)	5 (1.98)	8 (1.57)	14 (1.93)	16 (2.29)	0.639
Opioids	35 (8.91)	20 (7.91)	37 (7.27)	81 (11.17)	36 (5.15)	0.001
Antidepressants	64 (16.28)	50 (19.76)	99 (19.45)	137 (18.90)	107 (15.31)	0.214
Antiinfectives	92 (23.41)	66 (26.09)	111 (21.81)	154 (21.24)	98 (14.02)	<0.0001
Corticosteroids	61 (15.52)	58 (22.92)	110 (21.61)	145 (20.00)	79 (11.30)	<0.0001
Beta blocking agents	78 (19.85)	42 (16.60)	111 (21.81)	108 (14.90)	48 (6.87)	<0.0001
Anti-dementia drugs	29 (7.38)	17 (6.72)	48 (9.43)	62 (8.55)	62 (8.87)	0.669
Hypnotics and sedatives	23 (5.85)	13 (5.14)	41 (8.06)	59 (8.14)	38 (5.44)	0.133
Urological	38 (9.67)	17 (6.72)	54 (10.61)	52 (7.17)	53 (7.58)	0.133
Insulins and analogues	12 (3.05)	13 (5.14)	23 (4.52)	48 (6.62)	17 (2.43)	0.002

Figure 3:

AD progression subphenotypes. (a) Topic assignment trajectory in subphenotypes. The x-axis represents the time from MCI onset to AD onset. The y axis represents the topic assignment change in each subphenotype over time. Different colors represent different topics. (b) Survival curve for progression time. (c) Pairwise adjusted p-values for progression time between subphenotypes. (d) Stratified analysis of progression time on the whole cohort.

Subphenotype 1 (S1: diabetes and digestive system) consisted of 393 patients out of 2,579 patients who were diagnosed with both AD and MCI, and the MCI onset time was earlier than the AD onset time in the INSIGHT cohort. It was dominated by diabetes (T3) and digestive system (T26) related diseases. The median age within this cluster was 80, with females comprising 72.26% of the population. Compared to other subphenotypes, patients in this subphenotype demonstrated the second highest representation of Whites (43.76%). It also had the largest proportion of patients with gastroesophageal reflux disease (9.6%) and the smallest proportion of gastrointestinal agent usage (24.17%) at baseline. The disease trajectory exhibited consistent trends for the two predominant organ systems. Topic T3 (diabetes) showed a general increase over time, indicating a growing likelihood of an expanding number of patients being affected by diabetes. In contrast, Topic T26 (digestive diseases) displayed a generally high and stable trend over time, suggesting a consistent percentage of digestive diseases within the subphenotype. Interestingly, Topic T10 (sensory diseases) has a 1% probability of transitioning into Topic T3 (diabetes). Type 2 diabetes with ophthalmic manifestations, one of the top 10 diagnoses in T3, was characterized by high levels of blood glucose[36], which is associated with damage to the blood vessels in the eyes, leading to several eye-related complications such as diabetic retinopathy[36,37].

Subphenotype 2 (S2: endocrine and metabolic) included 253 patients. It was dominated by the endocrine and metabolic (T18), especially hypertension and endocrine related diseases. The median age within this cluster was 77, with females comprising 64.2% of the population. Compared to other subphenotypes, patients in this subphenotype have the largest proportion (31.2%) of longer progression time (>5 years). The disease trajectory exhibited a stable and consistent high pattern. The topic assignment of T18 manifested a consistently high and stable trend over time, suggesting a high and consistent percentage of endocrine and metabolic diseases within the subphenotype.

Subphenotype 3 (S3: cardiovascular) included 509 patients. It was dominated by cardiovascular system (T11) related diseases, especially heart rhythm disturbances and dizziness, as well as coronary atherosclerosis and cardiac pacemakers in situ. The median age of this subphenotype was 78 years. Additionally, patients in this subphenotype had the largest proportion of short progression time (36.34%, 1-2 years). It had the largest proportion of beta blocking agent usage (21.81%) and the largest proportion of heart disease (13.6%) at the baseline. The disease trajectory reflected a decreasing trend, that is, the topic assignment of T11 decreased over time, signifying that patients in this subphenotype were likely to “transfer” to non-cardiovascular or other diseases over time. There was a 3.5% probability of transitioning from Topic T11 to Topic T14, both of which are related to cardiovascular health. T14 is primarily dominated by heart failure issues. Heart rhythm disorders, associated with T11, could increase the risk of developing heart failure, as they can lead to hemodynamic instability and myocardial stress, thus contributing to the progression of heart failure[38].Additionally, there was a 1.5% probability of transitioning from T11 to T20, which is related to lipid-specific endocrine issues. Cholesterol is a major risk factor for the development of coronary heart disease, and chronic heart failure is associated with alterations in lipid metabolism, including changes in cholesterol and triglyceride levels[38,39].

Subphenotype 4 (S4: sensory and respiratory systems) included 725 patients. It was dominated by sensory system (T10) related diseases, such as glaucoma and cataracts, and respiratory diseases (T23). The median age was 79 and had the largest proportion of Blacks (16.13%). This subphenotype has the largest size compared to other subphenotypes. It has the second largest proportion of glaucoma and cataract cases at the baseline (6.76%). The sensory disease trajectory exhibited a stable high pattern, while the respiratory disease trajectory showed a decreasing pattern. The topic assignment for T10 showed a consistently high and stable trend over time, suggesting a consistent percentage of sensory abnormalities within the subphenotype. As the topic assignment of T23 decreased over time, the patients in this subphenotype were likely to transfer to non-respiratory diseases.

Subphenotype 5 (S5: musculoskeletal) included 699 patients and was dominated by musculoskeletal diseases (T1). Compared to the other subphenotypes, this subphenotype was the second largest and had the medium progression time (17.3%, 3-4 years). The disease trajectory exhibited a decreasing pattern. The topic assignment associated with T1 decreased over time, suggesting that patients within subphenotype 5 may “transfer out” to non-musculoskeletal diseases over time.

We observed distinct progression times from MCI to AD among the different subphenotypes (Figure 3b-c). Considering AD onset as an outcome event, the Kaplan-Meier curve demonstrated a significant difference in AD diagnosis rate among different subphenotypes (p value < 0.001) (Figure 3bc). Subphenotypes 1 and 2 had longer progression times compared to the other 3 subphenotypes, whereas subphenotype 3 (cardiac) had the shortest progression time. This suggested that individuals with digestive system problems are more likely to progress to AD at a slower rate, while those with cardiovascular issues may transition to AD in a relatively shorter time. Individuals with digestive system problems may have adopted healthier lifestyles, such as improved diet and stress management[40]. These changes can positively affect overall brain health and potentially slow the progression to AD[41]. Furthermore, digestive system problems may not directly influence metabolic pathways associated with AD, such as insulin resistance and cholesterol metabolism, to the same extent as other conditions like diabetes or cardiovascular diseases[41,42], although effects of the gastrointestinal microbiome on development of AD are still under intense evaluation.

For each subphenotype, we also have provided the sex-stratified and race-stratified analysis results with respect to patient demographics and progression time (Figure 3d). In the entire cohort, the progression time significantly differed among different sex and race groups and was also significantly different among different subphenotypes (Kruskal-Wallis test, p < 0.05). We further conducted sex-and race-stratified analyses within each subphenotype (Figure S4a), and the subphenotypes were not significantly different between males and females, or among the different (self-assigned) racial groups (Kruskal-Wallis test with Dunn's Multiple Comparison Test adjustment, p > 0.05). This suggested that our derived subphenotypes may have reduced the effects of sex and race on progression time.

We further investigated the trends of disease-related topics within different subphenotypes (Figure S5). Among the 13 coherent topics, 11 demonstrated a high trend in one unique subphenotype. For example, Topic T11, related to cardiovascular diseases, has the highest membership in Subphenotype 3, which was also dominated by cardiovascular-related issues. The trend for respiratory system diseases was the highest in Subphenotype 4, while the trend for musculoskeletal diseases was the highest in Subphenotype 5.

3.4. Early Detection of Subphenotypes Enhances Clinical Diagnostic Insights

We were able to detect AD progression subphenotypes using features extracted before MCI onset, which were not included during the initial subphenotype identification process. The logistic regression classifier was used to build the prediction model (Figure 4a). To ensure robustness, each cohort was randomly split into training and testing sets in an 80:20 ratio, and this process was repeated 10 times.

Figure 4:

Early prediction and cohorts’ concordance. (a) Pipeline of early prediction using before MCI onset features. (b) Accuracy of early prediction on 3 cohorts. (c) Similarities of coherent topics across the 3 cohorts. For each topic, we identified the top 10 diagnostic categories based on Phecode. Similarity was assessed by calculating the percentage of overlapping diagnostic categories. Additionally, we performed Fisher's exact test between pairs of topics and adjusted the p-values using the false discovery rate (FDR) method. Adjusted p-values less than 0.05, indicated by an asterisk, signify significant differences between the corresponding topics. Topics with non-significant adjusted p-values are considered similar across cohorts.

We achieved an accuracy over 0.7 in all three cohorts (Figure 4b). Specifically, using features collected within the last 6 months before MCI onset, we constructed prediction models that achieved accuracies of 0.72 ± 0.017, 0.71 ± 0.012, 0.86 ± 0.007 and AUCs of 0.90 ± 0.012, 0.91 ± 0.005, and 0.71 ± 0.017 in the OneFlorida, INSIGHT, UCSF cohorts, respectively. This indicates that our approach can predict patients MCI-to-AD progression trajectories well only relying on their health records collected within 6 months before onset of MIC. The identified subphenotypes captured the patterns of health condition progression from MCI to AD. Early prediction of a patient’s subphenotype membership may help clinicians predict his/her potential MCI-to AD progression trajectory and timing of MCI-to-AD transition, and make decisions and provide intervention early.

3.5. AD Progression subphenotypes were generalizable in independent databases

We repeated the same subphenotyping process shown in Figure 1 on the UCSF (Table S4) and OneFlorida+ (Table S5) cohorts. In each validation cohort we identified disease-associated coherent clinical topics, clustered subphenotypes related to specific organ systems, and demonstrated the potential trajectory of AD progression.

In the OneFlorida+ cohort, all learned potential topic patterns and subphenotypes are displayed in Figure S6 and Figure S7-S8. Here we see topics concentrated in conditions of the musculoskeletal system (T18), respiratory system (T28), cardiovascular system (T10/15), sensory organs (T16), mental disorders (T27), and genitourinary system (T22), where 8 out of 10 identified coherent topics were shared with the INSIGHT cohort (Figure 4c, Figure S6). In the subphenotypes, the identified 5 subphenotypes were dominated by endocrine and metabolic (S5), cardiovascular (S4), musculoskeletal (S2), respiratory systems (S1), and mental disorders (S3), which share four common results with the subphenotypes identified within the INSIGHT cohort. While INSIGHT has an unique digestive subphenotype, the OneFlorida+ cohort has an unique mental disorder subphenotype (S3). Subphenotype S3 had a longer progression time compared to the other 3 subphenotypes (Figure S7bc), whereas subphenotype S5 exhibited the shortest progression time. This indicated that individuals with mental disorders in OneFlorida+ may have had a slower progression to AD.

In the UCSF cohort, we have displayed all learned potential topics patterns and subphenotypes in Figure S9 and Figure S10-S11. We see topics concentrated in conditions of the musculoskeletal system (T6), respiratory system (T14), cardiovascular system (T1), sensory organs (T16) and topics dominated by endocrine and metabolic diseases, where 5 out of 5 identified coherent topics were the same as those in the INSIGHT cohort (Figure 4c). The 5 subphenotypes in the UCSF cohort were dominated by endocrine and metabolic (S1), cardiovascular (S3), musculoskeletal (S4), respiratory (S2), and sensory diseases (S5). These subphenotypes aligned with the five subphenotypes identified in the INSIGHT cohort. Notably, the subtype S4 in the INSIGHT cohort was a mixture of respiratory and sensory diseases, while in the UCSF cohort, S2 (respiratory) and S5 (sensory) were distinct subtypes, highlighting a difference in the disease structure between the two cohorts. UCSF had the lowest progression speed for the sensory diseases dominant subphenotype (S5) (Figure S10bc).

We also replicated the sex- and race-stratified analyses in both the UCSF and OneFlorida+ cohorts (Figure S7d, S10d). In both cohorts, the progression time was not significantly different among different race or sex groups in the entire cohort. However, the progression time was significantly different among different subphenotypes (p < 0.05). We conducted sex- and race-stratified analyses within each subphenotype, and the results were non-significant (p > 0.05, Figure S4bc). This suggested that our derived subphenotypes may have reduced the effects of age and race on progression time.

4. Discussion

We developed a data-driven framework to identify subphenotypes of AD progression based on EHR data. With the EHR from the INSIGHT cohort, we identified 5 subphenotypes dominated by musculoskeletal, cardiovascular, respiratory, digestive systems and endocrine/metabolic diseases. Across all cohorts, four of the five identified subphenotypes were highly consistent, whereas one subphenotype was more specific to individual cohorts, accounting for the variations observed between cohorts. Notably, the INSIGHT cohort exhibited a unique digestive disease-related subtype, the OneFlorida+ cohort had a unique mental disorder-related subtype, and the UCSF cohort presented a unique sensory subtype. The presence of unique subtypes within different cohorts suggests regional variations in the services/specialties offered at particular medical centers, variation in physician practice, differences in disease prevalence, or differences in patient lifestyle and environmental exposures. Even within subtypes dominated by the same organ system diseases, there were variations in the particular diseases. For instance, the musculoskeletal subtype in the OneFlorida+ cohort showed additional dominance with neurological diseases, indicating a potential overlap in the etiological pathways between musculoskeletal and neurological conditions. Indeed, disorders of the nervous system often result in musculoskeletal complaints: diabetic peripheral neuropathies, for instance, can lead to joint deformities at the foot and ankle[43,44].

Our analyses also revealed consistent coherent topics across cohorts, with 80% of coherent topics in the OneFlorida+ cohort and 100% of the INSIGHT cohort sharing the same organ system-dominant diseases. Different coherent topics across cohorts may share common non-dominant diseases. For example, the dermatology topic (T8) in the OneFlorida+ cohort shared common neoplastic diseases with the respiratory topic (T23) in the INSIGHT cohort. The coherent topics identified in one cohort may encompass diseases that are already represented within the topics of another cohort. For example, the OneFlorida+ cohort had a unique neurological system topic T25, which, although there was no corresponding neurological topic in INSIGHT, has a relatively high similarity with the T15 mental topic in INSIGHT, indicating potential overlaps in mental and neurological disease mechanisms[45,46]. Even when coherent topics were consistent across cohorts, they exhibited varying degrees of emphasis on specific diseases. For instance, the cardiovascular disease-related topics (T11,14,24) in the INSIGHT cohort were more focused on dizziness, heart failure, and heart issues, while in the OneFlorida+ cohort, the cardiovascular disease-related topics (T10,15) also encompassed cerebrovascular diseases like cerebral ischemia and cerebral artery occlusion.

Our study verified that subphenotypes such as respiratory, musculoskeletal, cardiovascular, and endocrine/metabolic could inform focused disease areas for treatment development for AD. For instance, managing respiratory disorders like chronic obstructive pulmonary disease (COPD) may help mitigate cognitive decline, as COPD has been linked to an increased risk of AD due to hypoxia and systemic inflammation[47-49]. Similarly, addressing musculoskeletal disorders and promoting physical activity could indirectly benefit cognitive health by reducing chronic pain and enhancing mobility[50-52]. Cardiovascular health is crucial for AD, as heart disease is a well-established risk factor for AD[53,54], with reduced cerebral blood flow and increased risk of cerebral embolism contributing to AD progression [55,56]. Moreover, controlling diabetes is vital, as insulin resistance and hyperglycemia are potential mechanisms linking diabetes to AD[57,58].

The cohort-specific subtypes–digestive, sensory, and mental disorders–also offer valuable insights. Digestive disorders may impact AD through the gut-brain axis, involving gut microbiota and inflammation[59,60]. Eye disorders such as glaucoma share pathological mechanisms with AD[61,62], and decreased vision is a risk factor for dementia, suggesting that early detection and management of eye conditions might also benefit these patients[63]. Moreover, sensory impairments such as hearing loss or vision problems may result in earlier detection of cognitive decline due to prompt medical attention[64,65]. Managing these impairments might indirectly slow the progression of Alzheimer’s disease. Additionally, mental disorders such as depression and anxiety are associated with an increased risk of AD[66] and managing these conditions might help slow AD progression[67,68]. Clustering AD patients into distinct subtypes and tailoring treatment accordingly could enhance precision medicine and personalized treatment strategies.

Of immediate clinical relevance is that the current work can accelerate the ability to identify slow or fast progressors, which could greatly accelerate AD clinical trials. AD clinical trials are hampered by significant inter-patient variability in disease progression, which dilutes the ability to see a drug’s effect and requires large sample sizes to demonstrate statistical significance[4,69]. To recruit and follow such large sample sizes, significant costs are incurred by funding agencies and pharmaceutical companies, which ultimately contribute to very high drug prices when a new drug is approved[69]. Thus, the current work offers a way to identify patients based on their estimated future rate of disease progression, which would lead to more efficiently run clinical trials, potentially lowering drug costs, and speeding the discovery of effective treatments for AD.

There is overlap between the subphenotypes and the topic pattern from the INSIGHT cohort. One topic pattern identified in INSIGHT was genitourinary. These conditions can be associated with sleep disruption due to the need to urinate and sleep disruption is an associated risk factor for progression to AD. Insufficient sleep has been associated with decreased clearance of amyloid from the brain via the glymphatics system.

Our study has several strengths. First, we addressed the issue of irregular visit intervals by aggregating records into three-month observation windows. Second, our DMTM design managed missing values by inferring missing events from neighboring ones (Supplementary Methods). Third, our framework enhances interpretability, which is crucial for making analysis results accessible and understandable to clinicians, by generating clinical topics that link organ systems to AD progression.

Our study also has limitations. First, the absence of biological data, such as biomarkers or genetic data, constraints biological interpretability. Second, varying cohort demographics could introduce noise into subphenotypes. Last, the lack of specificity of diagnoses can lead to incomplete or inaccurate records, potentially biasing findings.

However, our preprocessing steps, focusing on the transition from MCI to AD, offer some normalization, partially mitigating these issues.

5. Conclusion

In conclusion, our study dissects the complexity and heterogeneity of MCI to AD progression. These findings could be useful for clinicians and health systems in developing care models to meet the needs of patients with AD.

Statement of significance.

Problem or Issue	Alzheimer’s disease progression is highly complex, with patients showing diverse clinical manifestations. Identifying meaningful subphenotypes is critical for understanding disease mechanisms but is challenged by heterogeneous, sparse, high-dimensional data.
What is Already Known	Several data-driven approaches have been developed to identify and predict disease subphenotypes. However, these methods rarely account for the temporal evolution of clinical events or irregular visit intervals.
What this Paper Adds	The study introduces dynaPhenoM, a novel machine learning framework combining dynamic multimodal topic modeling and time-aware latent class analysis to identify and validate five clinically meaningful AD progression subphenotypes across three real-world EHR datasets, improving diagnostics and precision medicine strategies.

Data availability

The real-world data analyzed in this article were provided by OneFlorida+ Clinical Research Consortium (OneFlorida+ dataset), INSIGHT Clinical Research Network (INSIGHT dataset), and UCSF EHR database (UCSF dataset). These data are not publicly accessible due to restricted user agreement. Requests for access to OneFlorida+ dataset should be submitted to and approved by OneFlorida+ Clinical Research Consortium (https://www.ctsi.ufl.edu/ctsa-consortium-projects/oneflorida/); access to INSIGHT dataset can be obtained by contacting https://insightcrn.org/. The UCSF EHR database can be accessed by UCSF-affiliated individuals by contacting UCSF Clinical and Translational Science Institute (ctsi@ucsf.edu) or UCSF’s Information Commons team (info.commons@ucsf.edu). If the reader is unaffiliated with UCSF, they can set up an official collaboration with a UCSF-affiliated investigator by contacting the principal investigator, M.S.

Code availability

The implementation of the proposed DMTM and TLCA in Python and MATLAB are publicly available at https://github.com/manqizhou/AD-subtype. The data preprocessing steps were performed using the pandas, numpy, scipy, datetime, and tqdm packages in Python. The DMTM model was implemented using Python package Pydpm (3.0.1) (https://pypi.org/project/pydpm/). Evaluation metrics were computed using the sklearn.metrics package in Python.