CoChat

Unsupervised Deep Learning of Electronic Health Records to Characterize Heterogeneity Across Alzheimer Disease and Related Dementias: Cross-Sectional Study

Open access
West, Matthew, Cheng, You, He, Yingnan, Leng, Yu, Magdamo, Colin, Hyman, Bradley T, Dickson, John R, Serrano-Pozo, Alberto, Blacker, Deborah, Das, Sudeshna · 2025 · JMIR Aging

Published version of a preprint

The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index.

Abstract

Background: Alzheimer disease and related dementias (ADRD) exhibit prominent heterogeneity. Identifying clinically meaningful ADRD subtypes is essential for tailoring treatments to specific patient phenotypes. Objective: We aimed to use unsupervised learning techniques on electronic health records (EHRs) from memory clinic patients to identify ADRD subtypes. Methods: We used pretrained embeddings of non-ADRD diagnosis codes (International Classification of Diseases, Ninth Revision) and large language model (LLM)–derived embeddings of clinical notes from patient EHRs. Hierarchical clustering of these embeddings was used to identify ADRD subtypes. Clusters were characterized regarding their demographic and clinical features. Results: We analyzed a cohort of 3454 patients with ADRD from a memory clinic at Massachusetts General Hospital, each with a specialist diagnosis. Clustering pretrained embeddings of the non-ADRD diagnosis codes in patient EHRs revealed the following 3 patient subtypes: one with skin conditions, another with psychiatric disorders and an earlier age of onset, and a third with diabetes complications. Similarly, using LLM-derived embeddings of clinical notes, we identified 3 subtypes of patients as follows: one with psychiatric manifestations and higher prevalence of female participants (prevalence ratio: 1.59), another with cardiovascular and motor problems and higher prevalence of male participants (prevalence ratio: 1.75), and a third one with geriatric health disorders. Notably, we observed significant overlap between clusters from both data modalities (χ24=89.4; P<.001). Conclusions: By integrating International Classification of Diseases, Ninth Revision codes and LLM-derived embeddings, our analysis delineated 2 distinct ADRD subtypes with sex-specific comorbid and clinical presentations, offering insights for potential precision medicine approaches.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.2196/65178

Cites

62 of 62 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Visualization of the clustering pipeline for (A) International Classification of Diseases (ICD) codes and (B) notes. For each subfigure, the workflow goes from left to right. BERT: Bidirectional Encoder Representations from Transformers.
Figure 1 Visualization of the clustering pipeline for (A) International Classification of Diseases (ICD) codes and (B) notes. For each subfigure, the workflow goes from left to right. BERT: Bidirectional Encoder Representations from Transformers.CC BY · PubMed Central
Figure 2 CONSORT (Consolidated Standards of Reporting Trials) diagram illustrating the selection of patients from the Massachusetts General Hospital (MGH) electronic health record (EHR) system. ADRD: Alzheimer disease and related dementias; Dx: diagnosis; ICD: International Classification of Diseases.
Figure 2 CONSORT (Consolidated Standards of Reporting Trials) diagram illustrating the selection of patients from the Massachusetts General Hospital (MGH) electronic health record (EHR) system. ADRD: Alzheimer disease and related dementias; Dx: diagnosis; ICD: International Classification of Diseases.CC BY · PubMed Central
Figure 3 Heatmap of enrichment in International Classification of Diseases (ICD) clusters and topics in note clusters. (A) This heatmap displays the enrichment of ICD-9 codes across ICD embedding clusters. Cluster 1 is primarily dominated by skin-related and certain cardiovascular conditions. Cluster 2 is marked by its exclusive and high prevalence ratios (PR) in psychiatric and behavioral conditions. Cluster 3 shows a diverse set of conditions with a significant prevalence of respiratory, pain-related, and complicated diabetic mellitus. (B) This heatmap displays representative words for each note cluster identified through topic modeling. Cluster 1 is primarily dominated by psychiatric manifestations and medications. Cluster 2 highlights cardiovascular, motor, and sensory issues. Cluster 3 covers a variety of symptoms and conditions, including autoimmune issues, behavioral and movement disorders, sleep disturbances, etc. In both (A) and (B), the color intensity of each code or word-cluster pairing reflects the PR (prevalence of code or word in observed group divided by prevalence in other groups) associated with that code or word. Words colored as exclusive were only present in one cluster.
Figure 3 Heatmap of enrichment in International Classification of Diseases (ICD) clusters and topics in note clusters. (A) This heatmap displays the enrichment of ICD-9 codes across ICD embedding clusters. Cluster 1 is primarily dominated by skin-related and certain cardiovascular conditions. Cluster 2 is marked by its exclusive and high prevalence ratios (PR) in psychiatric and behavioral conditions. Cluster 3 shows a diverse set of conditions with a significant prevalence of respiratory, pain-related, and complicated diabetic mellitus. (B) This heatmap displays representative words for each note cluster identified through topic modeling. Cluster 1 is primarily dominated by psychiatric manifestations and medications. Cluster 2 highlights cardiovascular, motor, and sensory issues. Cluster 3 covers a variety of symptoms and conditions, including autoimmune issues, behavioral and movement disorders, sleep disturbances, etc. In both (A) and (B), the color intensity of each code or word-cluster pairing reflects the PR (prevalence of code or word in observed group divided by prevalence in other groups) associated with that code or word. Words colored as exclusive were only present in one cluster.CC BY · PubMed Central
Figure 4 Clustering of International Classification of Diseases (ICD) embeddings and their demographic and diagnostic associations. (A) Uniform Manifold Approximation and Projection (UMAP) visualization of ICD embeddings were characterized by 3 clusters: cluster 1 includes 1501 (43.5%) patients, cluster 2 comprises 1597 (46.2%) patients, and cluster 3 contains 356 (10.3%) patients. (B) Bar plot showing prevalence for sex by cluster, significance based on a chi-square test. Notably, cluster 3 has a significantly higher proportion of female participants compared to male participants (PFDR=.009). (C) Violin plot illustrating the distribution of age of onset across clusters. Each violin plot shows the kernel density estimate of the data, with the center line representing the median age of onset. Box plot elements are overlaid, where the box limits indicate the upper and lower quartiles, and the whiskers extend to 1.5 times the IQR. Individual points are hidden for clarity. Significant differences are observed, with cluster 2 showing the earliest average age of onset at 69.8 (SD 9.6) years, and cluster 3 the latest at 75.5 (SD 8.1) years (P<.001). (D) Heatmap showing prevalence ratio for Alzheimer disease and related dementias (ADRD) diagnoses across clusters, significance derived from a chi-squared test. Clinical diagnoses include Alzheimer disease (AD); dementia unspecified; frontotemporal dementia (FTD); Lewy body dementia (LBD); vascular cognitive impairment (VCI); and others such as posterior cortical atrophy (PCA), progressive supranuclear palsy (PSP), corticobasal degeneration (CBD), and primary progressive aphasia (PPA). Significant distribution variations are evident across clusters. The circle size, color, and number indicate the magnitude of the prevalence ratio. Blue indicates underrepresentation while red indicates overrepresentation. P values were corrected with false discovery rate adjustments.
Figure 4 Clustering of International Classification of Diseases (ICD) embeddings and their demographic and diagnostic associations. (A) Uniform Manifold Approximation and Projection (UMAP) visualization of ICD embeddings were characterized by 3 clusters: cluster 1 includes 1501 (43.5%) patients, cluster 2 comprises 1597 (46.2%) patients, and cluster 3 contains 356 (10.3%) patients. (B) Bar plot showing prevalence for sex by cluster, significance based on a chi-square test. Notably, cluster 3 has a significantly higher proportion of female participants compared to male participants (PFDR=.009). (C) Violin plot illustrating the distribution of age of onset across clusters. Each violin plot shows the kernel density estimate of the data, with the center line representing the median age of onset. Box plot elements are overlaid, where the box limits indicate the upper and lower quartiles, and the whiskers extend to 1.5 times the IQR. Individual points are hidden for clarity. Significant differences are observed, with cluster 2 showing the earliest average age of onset at 69.8 (SD 9.6) years, and cluster 3 the latest at 75.5 (SD 8.1) years (P<.001). (D) Heatmap showing prevalence ratio for Alzheimer disease and related dementias (ADRD) diagnoses across clusters, significance derived from a chi-squared test. Clinical diagnoses include Alzheimer disease (AD); dementia unspecified; frontotemporal dementia (FTD); Lewy body dementia (LBD); vascular cognitive impairment (VCI); and others such as posterior cortical atrophy (PCA), progressive supranuclear palsy (PSP), corticobasal degeneration (CBD), and primary progressive aphasia (PPA). Significant distribution variations are evident across clusters. The circle size, color, and number indicate the magnitude of the prevalence ratio. Blue indicates underrepresentation while red indicates overrepresentation. P values were corrected with false discovery rate adjustments.CC BY · PubMed Central
Figure 5 Clustering of note embeddings and their demographic and Alzheimer disease and related dementias (ADRD) diagnosis associations. (A) Uniform Manifold Approximation and Projection (UMAP) visualization of note embeddings characterized by 3 clusters: cluster 1 includes 1280 (37.1%) patients, cluster 2 includes 1161 (33.6%) patients, and cluster 3 includes 1013 (29.3%) patients. (B) Bar plot showing prevalence for sex by cluster, with significance based on chi-square tests. Notably, cluster 1 and cluster 3 were both enriched by female participants (P<.001) while cluster 2 was enriched by male participants (P<.001). (C) Violin plot illustrating the distribution of age of onset across clusters. Each violin plot shows the kernel density estimate of the data, with the center line representing the median age of onset. Box plot elements are overlaid, where the box limits indicate the upper and lower quartiles, and the whiskers extend to 1.5 times the IQR. Individual points are hidden for clarity. Significant differences are observed, with cluster 2 showing the latest average age of onset at 72.8 (SD 9.2) years, and cluster 1 the earliest at 71.5 (SD 9.3) years (P<.001). (D) Heatmap showing prevalence ratio for ADRD diagnoses across clusters, with significance derived from a chi-square test. Diagnoses include Alzheimer disease (AD); dementia unspecified; frontotemporal dementia (FTD); Lewy body dementia (LBD); vascular cognitive impairment (VCI); and others such as posterior cortical atrophy (PCA), progressive supranuclear palsy (PSP), corticobasal degeneration (CBD), and primary progressive aphasia (PPA). No significant distribution variations are observed across clusters (P>.05). The circle size, color, and number indicate the magnitude of the prevalence ratio. Blue indicates underrepresentation while red indicates overrepresentation. P values were corrected with false discovery rate (FDR) adjustments.
Figure 5 Clustering of note embeddings and their demographic and Alzheimer disease and related dementias (ADRD) diagnosis associations. (A) Uniform Manifold Approximation and Projection (UMAP) visualization of note embeddings characterized by 3 clusters: cluster 1 includes 1280 (37.1%) patients, cluster 2 includes 1161 (33.6%) patients, and cluster 3 includes 1013 (29.3%) patients. (B) Bar plot showing prevalence for sex by cluster, with significance based on chi-square tests. Notably, cluster 1 and cluster 3 were both enriched by female participants (P<.001) while cluster 2 was enriched by male participants (P<.001). (C) Violin plot illustrating the distribution of age of onset across clusters. Each violin plot shows the kernel density estimate of the data, with the center line representing the median age of onset. Box plot elements are overlaid, where the box limits indicate the upper and lower quartiles, and the whiskers extend to 1.5 times the IQR. Individual points are hidden for clarity. Significant differences are observed, with cluster 2 showing the latest average age of onset at 72.8 (SD 9.2) years, and cluster 1 the earliest at 71.5 (SD 9.3) years (P<.001). (D) Heatmap showing prevalence ratio for ADRD diagnoses across clusters, with significance derived from a chi-square test. Diagnoses include Alzheimer disease (AD); dementia unspecified; frontotemporal dementia (FTD); Lewy body dementia (LBD); vascular cognitive impairment (VCI); and others such as posterior cortical atrophy (PCA), progressive supranuclear palsy (PSP), corticobasal degeneration (CBD), and primary progressive aphasia (PPA). No significant distribution variations are observed across clusters (P>.05). The circle size, color, and number indicate the magnitude of the prevalence ratio. Blue indicates underrepresentation while red indicates overrepresentation. P values were corrected with false discovery rate (FDR) adjustments.CC BY · PubMed Central
Figure 6 Heatmap of the association of International Classification of Diseases (ICD) clusters with note clusters. Heatmap of the association of ICD clusters with note clusters. ICD clusters were significantly associated with note clusters (P<.001). Post hoc analyses revealed that note cluster 1 was positively associated with ICD cluster 2 (standardized residual=8.42, P<.001) and negatively associated with ICD clusters 1 and 3 (cluster 1: standardized residual=−4.57, P<.001; cluster 3: standardized residual=−6.36, P<.001). Furthermore, note cluster 2 was positively associated with ICD cluster 3 (standardized residual=4.9, P<.001) and negatively associated with ICD cluster 2 (standardized residual=−5.48, P<.001). Finally, note cluster 3 was negatively associated with ICD cluster 2 (standardized residual=−3.25, P=.003). Each cell displays standardized residuals (standardized differences between the observed count and the expected count) along with the count of overlapping patients (n). Color bar represents the value of the standardized residual. P values were corrected with false discovery rate adjustments.
Figure 6 Heatmap of the association of International Classification of Diseases (ICD) clusters with note clusters. Heatmap of the association of ICD clusters with note clusters. ICD clusters were significantly associated with note clusters (P<.001). Post hoc analyses revealed that note cluster 1 was positively associated with ICD cluster 2 (standardized residual=8.42, P<.001) and negatively associated with ICD clusters 1 and 3 (cluster 1: standardized residual=−4.57, P<.001; cluster 3: standardized residual=−6.36, P<.001). Furthermore, note cluster 2 was positively associated with ICD cluster 3 (standardized residual=4.9, P<.001) and negatively associated with ICD cluster 2 (standardized residual=−5.48, P<.001). Finally, note cluster 3 was negatively associated with ICD cluster 2 (standardized residual=−3.25, P=.003). Each cell displays standardized residuals (standardized differences between the observed count and the expected count) along with the count of overlapping patients (n). Color bar represents the value of the standardized residual. P values were corrected with false discovery rate adjustments.CC BY · PubMed Central
Figure 9 Inline graphic 3.
Figure 9 Inline graphic 3.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 62 references

Research works printed but not linked: 9

  1. He, Z, 2022, AMIA Jt Summits Transl Sci Proc, 2022, 226
  2. Vaswani, A, arXiv
  3. DevlinJChangMWLeeKToutanovaKBERT: pre-training of deep bidirectional transformers for language understandingProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies2019NAACL-HLT '19June 2-7, 2019Minneapolis, MN417186
  4. Choi, Y, 2016, AMIA Jt Summits Transl Sci Proc, 2016, 41
  5. Grootendorst, M, arXiv
  6. Achiam, OJ, arXiv
  7. EsterMKriegelHPSanderJXuXA density-based algorithm for discovering clusters in large spatial databases with noiseProceedings of the Second International Conference on Knowledge Discovery and Data Mining1996KDD '96August 2-4, 1996Portland, OR22631
  8. RibinoPNapoliCDParagliolaGSerinoLGaspariniFChiccoDExploratory analysis of longitudinal data of patients with dementia through unsupervised techniquesProceedings of the 4th Italian Workshop on Artificial Intelligence for an Ageing Society2023AIxAS '23November 6-9, 2023Rome, Italy121
  9. Eyre, H, 2021, AMIA Annu Symp Proc, 2021, 438

Cited by

Metadata

Authors
West, M., Cheng, Y., He, Y., Leng, Y., Magdamo, C., Hyman, B. T., Dickson, J. R., Serrano-Pozo, A., Blacker, D., & Das, S.
Affiliations
Massachusetts General Hospital (US)
Corresponding author
Das, Sudeshna
Venue
JMIR Aging
Published
2025-03-31
Volume
8
Pages
e65178
ISSN
2561-7605 (ISSN-L)
OpenAlex
W4409003461
Cited by
8
References
59
Citations by year
2026: 4 · 2025: 4
FWCI
3.20
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • JMIR Aging (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Institute on Aging (P30AG062421)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Psychiatry and Mental health › Dementia and Cognitive Impairment Research
  • Health Sciences › Health Professions › Health Information Management › Medical Coding and Health Information
Keywords
Alzheimer's disease and related dementias, dementia subtypes, patient heterogeneity, electronic health records, large language model embeddings, clinical notes, memory clinic patients, psychiatric comorbidity, sex differences, diabetes complications, cardiovascular comorbidity, precision medicine
MeSH terms
Unsupervised Machine Learning, Deep Learning, Aged, Aged, 80 and over, Alzheimer Disease, Cross-Sectional Studies, Dementia, Female, Humans, Male, Middle Aged, Electronic Health Records

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.