Stratifying heart failure patients with graph neural network and transformer using Electronic Health Records to optimize drug response prediction

Abstract

Objectives

Heart failure (HF) impacts millions of patients worldwide, yet the variability in treatment responses remains a major challenge for healthcare professionals. The current treatment strategies, largely derived from population based evidence, often fail to consider the unique characteristics of individual patients, resulting in suboptimal outcomes. This study aims to develop computational models that are patient-specific in predicting treatment outcomes, by utilizing a large Electronic Health Records (EHR) database. The goal is to improve drug response predictions by identifying specific HF patient subgroups that are likely to benefit from existing HF medications.

Materials and Methods

A novel, graph-based model capable of predicting treatment responses, combining Graph Neural Network and Transformer was developed. This method differs from conventional approaches by transforming a patient's EHR data into a graph structure. By defining patient subgroups based on this representation via K-Means Clustering, we were able to enhance the performance of drug response predictions.

Results

Leveraging EHR data from 11 627 Mayo Clinic HF patients, our model significantly outperformed traditional models in predicting drug response using NT-proBNP as a HF biomarker across five medication categories (best RMSE of 0.0043). Four distinct patient subgroups were identified with differential characteristics and outcomes, demonstrating superior predictive capabilities over existing HF subtypes (best mean RMSE of 0.0032).

Discussion

These results highlight the power of graph-based modeling of EHR in improving HF treatment strategies. The stratification of patients sheds light on particular patient segments that could benefit more significantly from tailored response predictions.

Conclusions

Longitudinal EHR data have the potential to enhance personalized prognostic predictions through the application of graph-based AI techniques.

Introduction

Heart failure (HF) has rapidly evolved into a global epidemic, impacting an estimated 64 million individuals globally.¹ Far from being a singular disease, HF presents as a complex clinical syndrome with profound implications for morbidity, mortality, and healthcare expenditures. The multifaceted nature of HF, stemming from its varied origins and pathologies, gives rise to a spectrum of patient subgroups. This diversity complicates treatment, as the efficacy of interventions varies across patients.² Addressing this challenge necessitates a deeper understanding of the unique traits characterizing these diverse phenotypes, which would aid in both diagnosis and treatment monitoring. However, prevailing treatments, rooted in randomized controlled trials, fall short in catering to this phenotypic diversity, often leading to biases and restricted applicability.³

The variability in drug responses can be attributed to differences in patients' clinical and biomarker profiles. This underscores the potential of precision medicine, which emphasizes tailored treatments, as a promising avenue for HF management. While the spotlight in precision medicine often shines on “big data” from omics research, the limited genetic facets of HF have redirected focus towards Electronic Health Records (EHRs).⁴ These records house vast amounts of time-sensitive clinical data, capturing the rich phenotypic diversity that reflects varied clinical and biomarker profiles.⁵ Traditional deep learning models (Supplementary Note S1) have demonstrated potential in forecasting treatment outcomes.^6–8 However, existing computational methods struggle to capture the intricate topological structure of non-euclidean data.⁹ The rich tapestry of clinical events in EHRs, combined with the temporal progression and phenotypical dynamics (known as heterogeneity) of HF,¹⁰ makes them ripe for graph-based representations. Yet, existing graph representation techniques tend to prioritize temporal patterns, overlooking the spatial intricacies of the phenotypical features.¹¹

This study addresses these gaps by introducing an innovative framework that combines the strengths of Graph Neural Network (GNN) and Transformer technologies. This framework is designed to achieve more precise outcome predictions based on a stratification of patients using phenotypical representations, which were captured from both the temporal dynamics and the heterogeneity presentation in EHR data, such as the variability and changing correlations among phenotypical features over time. This is achieved by transforming longitudinal patient data into concept graphs and employing a GNN model guided by the Transformer’s Self-Attention mechanism to learn patient graph representations that are optimized for drug response prediction. With the application of unsupervised cluster analysis (ie, K-Means Clustering) to the learned patient graph representations, we were subsequently able to derive phenotypically distinct patient subgroups (ie, phenogroups). Utilizing a comprehensive dataset from the Mayo Clinic that includes data on 11 627 heart failure (HF) patients, our method has shown exceptional effectiveness in predicting responses to drugs impacting the N-terminal pro-B-type natriuretic peptide (NT-proBNP),¹² a biomarker critical for evaluating both short-term and long-term cardiovascular outcomes.¹³ The model’s performance has been markedly superior across various benchmarks, outperforming traditional models and the rudimentary HF categorizations.

Methods

Study population

The data for the study were obtained from Mayo Clinic’s United Data Platform (UDP), a data warehouse that contains, consolidates, and standardizes all clinical data collected within the institution. We initially identified 46 065 HF patients using the diagnosis codes listed in Supplementary Table S1. With each patient record corresponding to a single visit, we utilized the demographics, diagnosis, lab test, and medication information from March 1986 until February 2023. In the selection process (Supplementary Figure S1), we included patients who were above 18 years old age and had at least one NT-proBNP lab assessment. We further filtered out any patients who had any erroneous measurements in their EHR records. Based on these inclusion criteria, the final cohort had 11 627 HF patients in total. We evaluated the pharmacological effect of five categories/classes of drugs: Angiotensin-converting-enzyme inhibitor (ACEI), β-blocker (BB), Angiotensin II receptor blocker (ARB), Statin and Loop Diuretic (LD). Among these, ACEI, BB, ARB, and LD are recognized as guideline-directed medical therapy (GDMT) for heart failure.¹⁴ Although Statins are usually not considered as GDMT, given their widespread use among heart failure patients and evidence from previous studies showing they could reduce NT-proBNP level^15,16—a primary focus of this study—we also included Statins in our analysis. Subsequently, the study cohort for each drug class was created separately by using the corresponding medication codes listed in Supplementary Table S2. The detailed data statistics per drug class for the study cohorts are shown in Table 1. We use N_pat in the rest of the article to denote the total number of patients in each dataset. The use of data for this study was approved by the Mayo Clinic Institutional Review Board.

Table 1.

Summary table of data statistics per drug class.

	ACEI	BB	ARB	STATIN	LD
Num. patients in total, Npat	2214	2762	2766	6298	5674
Avg. visits per patient	2	2	2	2	2
Gender, n (%)
Female	929 (42)	1276 (46)	1405 (51)	2839 (45)	2845 (50)
Male	1285 (58)	1486 (54)	1361 (49)	3459 (55)	2829 (50)
Average age (SD)	77.93 (12.04)	78.98 (12.14)	77.82 (12.06)	78.03 (11.84)	78.53 (12.08)
Race and ethnicity, n (%)
African American	23 (1)	28 (1)	75 (2.7)	128 (2)	48 (0.8)
Hispanic or Latino	4 (0.4)	6 (0.2)	8 (0.3)	20 (0.3)	15 (0.2)
Non-Hispanic White	2129 (96)	2650 (96)	2568 (93)	5938 (94)	5488 (97)
Others	58 (2.6)	78 (2.8)	115 (4)	212 (3.7)	123 (2)

Problem definition

In this retrospective observational study, we learn patient graph representations based on the longitudinal patient history in EHR. The patient history can be perceived as a collection of EHR records associated with HF conditions, where each record represents a time-ordered visit and each visit is comprised of a list of clinical concepts essentially summarizing the prognostic and interventional events involved in the HF management. Formally, let P = (V₁, V₂, …., V_T) denote the EHR records of a single patient with total T visits, where V_i = (c₁ⁱ, c₂ⁱ, …., c_|v_i|ⁱ) is a visit in P arranged by the time of occurrence and ${\mathit{c}}_{\mathrm{j}}^{\mathrm{i}}$ = (${\mathit{e}}_{\mathrm{j}}^{\mathrm{i}}$, ${\mathit{v}}_{\mathrm{j}}^{\mathrm{i}}$, ${\mathit{t}}_{\mathrm{j}}^{\mathrm{i}}$) is a clinical event in V_i composed of a tuple of the type of event ${\mathit{e}}_{\mathrm{j}}^{\mathrm{i}}$ ϵ E, the observed value of the event ${\mathit{v}}_{\mathrm{j}}^{\mathrm{i}}$ and the timestamp of the event ${\mathit{t}}_{\mathrm{j}}^{\mathrm{i}}$ ϵ R^*₊. Here, E corresponds to the unique set of clinical events, ${\mathit{v}}_{\mathrm{j}}^{\mathrm{i}}$ is either a categorical value or a numerical measurement depending on the type of event and R^*₊ is the set of positive real numbers. Given the patient’s EHR sequence P containing the time-varying heterogeneous phenotypic events from E, a concept graph G is constructed from P and is optimized for the prediction of the patient’s drug response, ŷ ϵ [0,1], in the future time step (ie, last visit), as measured by the biomarker (ie, normalized NT-proBNP lab test), via learning a graph-based mapping function ƒ: P → ŷ.

Graph construction

To model the process of patient's treatment in the clinical setting, we formulated a health network for each individual by converting patient-specific events from the EHR sequence P into a concept graph G = (V, E, A, X), where V represents vertices, E the connecting edges, A the adjacency matrix, and X the node feature matrix. Nodes V are derived from three EHR data sources: Demographics (age, sex), Laboratory tests (including hemodynamic variables diastolic blood pressure (DBP), systolic blood pressure (SBP), and systematic vascular resistance (SVR)), and Diagnosis Comorbidity (encompassing six prevalent conditions like hypertension and diabetes mellitus). We explored graph construction from single and multi-data source perspectives, using each data source independently in the former and combining variables from all three sources in the latter. Our graph construction is motivated by the EHR data structure, which can be decomposed into the temporal and spatial dependencies.¹⁷ As shown in Figure 1 step 2 (“Concept Graph Construction”), each variable (eg, DBP) in the patient’s EHR can be assessed over multiple visits such that any future physiological changes are dependent on the values recorded in the previous visits. This time-related dependence of a variable with itself over sequential visits is regarded as temporal correlation and reflects the patient’s HF progression over time with respect to that variable. Also, for a specific visit, different types of variables (known as heterogeneity) can have influence on each other. This kind of latent dependencies is known as spatial/heterogeneity correlations and defines the possible physiological interactions among different variables over the course of HF treatment. These temporal-heterogeneity correlations indicate how different variables relate to each other over different time steps (visits) and thus exist in a noneuclidean space. As such, nodes are defined considering both the event type (heterogeneity) and its timing (temporal), with future events within a time window T = 5 treated as one-hop neighbors to capture long-term dependencies. The adjacency matrix A ϵ R^{|V| x |V|} summarizes this temporal-heterogeneity graph structure knowledge whereby its (i, j)-th entry is 1 if e(i, j) ϵ E, otherwise it is a 0, as exemplified in the following example. For a given patient P = (V₁, V₂,….V_T) with T medical visits, we will transform the medical concepts from each visit into nodes within A, with each matrix entry corresponding to a medical concept from a visit. For example, for a diagnosis “hypertension” will be represented as T nodes as each hypertension₁, hypertension₂,… hypertension_T. To record the temporal and heterogeneity relation to represent the progression, we will link all the medical events within a time window (ie, T = 5). For example, if hypertension is diagnosed in visit 1 and DBP is assessed in visit 2, then the value will be set to 1 for the entry corresponding to (hypertension₁, DBP₂). Although Demographics node, such as age, per se has a predictable longitudinal property, it is important to note that this information is not used in all the graph constructions. We included Demographics nodes in the multi-data source graph construction to investigate how its hidden temporal-heterogeneity correlations with concepts from other data sources (ie, Lab, Comorbidity) would impact performance. We employed an undirected graph topology to facilitate bidirectional information flow of past and future events, a decision backed by preliminary results and previous studies.¹⁸ Since our base model, the Transformer is inherently sequence-order invariant,¹⁹ we utilized the event's timestamp to inject positional information into the graph structure, ensuring a comprehensive and nuanced representation of the patient’s healthcare journey. Refer to Supplementary Note S2 for the graph construction in details.

Figure 1.

Proposed graph-based framework to leverage temporal heterogeneity relations in the EHR. (1) We pass as input the longitudinal observations associated with three data sources in EHR (ie, lab test, demographics, and comorbidity), which exist as a multivariable event sequence. The study cohorts of HF patients are categorized based on five drug classes—ACEI, β-blocker (BB), ARB, Statin or loop diuretic (LD). (2) The temporal-heterogeneity correlations are latent relationships that span across different visits (temporal dimension) between different variables, where each visit could contain multiple variables, such as age, diagnosis codes, and lab tests, recorded over different time points over the visit duration. Therefore, each variable is connected to itself over different visits through temporal correlations and connected to itself over different visits through temporal-heterogeneity correlations. (3) The proposed model architecture consists of the Graph Transformer component augmented with the global self-attention mask and followed by a GNN layer. For the input graph shown on the lower left, learning the representation of node 0 (highlighted) would input the feature vectors of the sampled neighbors, nodes 0, 1, 2, and 3, into the Graph Transformer component for the computation of the self-attention score. The global attention mask is incorporated to ensure that the neighbors are selected based on the co-occurrence of events. Note that the node numbering is only for indexing purposes and does not indicate the positional information. (4) K-Means Clustering is applied to stratify patients into phenogroups and is reproduced with validation cohorts. We evaluate the model’s performance against traditional models and investigate the important predictors. In addition, the performance of the identified phenogroup cohorts are compared against the traditional HF subtypes.

Drug response prediction

As Figure 1 (step 3) shows, we used a GNN model to map the concept graph to a low-dimensional vector. Given the input feature matrix X of all the nodes in G, we used self-attention to project X to the query, key, and value spaces, Q = XWq, K = XWk, V = XWv, using the trainable weight matrices Wq, Wk, and Wv, respectively. Then scaled dot product is employed to compute the attention as Z = Softmax$\left(\frac{\mathrm{Q}{K}^T}{\surd d}\right)$ V, where d is the dimensionality of the self-attention that is used in the scaling factor for numerical stability, named as Graph Transformer in this article. We adopted a sampling strategy²⁰ that samples a subset of the nodes as the neighborhood, Ɲn, for each node and feeds it as the input matrix into the self-attention component. Further to enable the Graph Transformer to process the temporal-heterogeneity relations in the input graph, we augment a GNN layer on top of the Graph Transformer as Z′ = GNN(A, Z), where A is the adjacency matrix. We also apply a global attention mask to the patient graphs with prognostic patterns. Specifically, we drew a binary event co-occurrence matrix, M ϵ R^{|V| x |V|} from all the patient's records in the EHR. We then use this event co-occurrence matrix to redefine self-attention with the attention mask function Mask, notated as, Z = Softmax$\left(\frac{\mathit{Mask}\left(\mathrm{Q}K\right)}{\surd d}\right)$ V, $\mathit{Mask}\left(\mathrm{QK}\right)=\left\{\begin{array}{c}1, \mathrm{M}[\mathrm{i},\mathrm{ }\mathrm{j}]\mathrm{ }=\mathrm{ }1 \\ -\mathrm{\infty },\mathrm{ }\mathrm{otherwise}\end{array}\right.$, where i and j are the positions in the query and key, respectively. Finally, the output node representation Z′ is projected into a single vector ẑ and passed through a linear layer to forecast the NT-proBNP measurement ŷ. Refer to Supplementary Notes S3 and S4 for the model and training details, respectively.

Patient stratification

After optimizing our Graph-Transformer model on drug response prediction using the EHR-derived patient graphs as inputs, the graph-based deep patient representations (ẑ) learned by Graph-Transformer are then subjective to an unsupervised clustering approach (K-Means) for patient stratification into phenotypically distinct subgroups with the potential to provide therapeutic insights. Phenomapping^21–23 has emerged as an unsupervised technique that, instead of making an assumption on the number of clusters based on a priori knowledge, analyzes the phenotypic patient data for association patterns to identify distinct patient subgroups (ie, phenogroups). We performed phenomapping by applying K-means clustering to the learned patient graph representations of all 11627 HF patients in our study population. The optimal number of clusters K is determined by the Davies-Bouldin Index (DBI)²⁴ and the Silhouette Score (SS)²⁵ (See Supplementary Note S5 for the details of determination of cluster numbers); the clustering results with K = 4 are visualized in Figure 2A.

Figure 2.

Quantitative performance comparisons (A) with different data sources in EHR as the input, (B) against baseline models, (C) against ablated model variants, and (D) single therapy vs combination therapy.

We performed independent validation analyses on two subsets consisting of 6000 randomly sampled patients from the 11627 to verify that the original phenomapping can be replicated on smaller cohorts. We set K to the same number estimated from the original cohort and performed K-Means clustering on the two validation cohorts separately.

Results

Model evaluation

We compared the proposed model’s (Graph-Transformer) performance in predicting drug response against previous drug response models as well as other deep learning baseline methods. Additionally, we characterized the performance trade-offs of the proposed model with its ablated variants and data source for ACEI medications in root mean squared error (RMSE) in Figure 2.

Graph-Transformer performed better on the single data source Lab than the multi-data source (ALL), with a performance gap of 13%, although the difference is not statistically significant (ρ = 0.09 > 0.05) (Figure 2A). Comparing the proposed model’s performance with other baseline models in Figure 2B, (please refer to the Supplementary Note S6 for their implementation details), our model consistently outperformed all the baselines significantly (ρ = 0.001). The ablation study in Figure 2C suggests that the proposed model achieved 23% (ρ = 0.02) and 9% (ρ = 0.42) RMSE reductions over its individual components, GNN and Graph Transformer, respectively.

Effect of drugs and combinations

As recommended by the international guidelines for HF management, typically a combination therapeutic approach involving drugs across multiple drug classes is prescribed to the generally elderly HF population with multimorbidity. In light of this, we analyze the treatment response by tapping into the medication information from five different angles: (1) HF drug class with single vs combination therapy. In single therapy, the input cohort includes patients who are taking medication/s from only one drug class exclusively before the study end period. On the other hand, for combination therapy, patients taking medications across multiple drug classes before the study end period are included in the input cohort; (2) top drug per class assesses the drug response prediction performance of the most frequently taken drug in each drug class; input cohort per class includes patients who are taking the respective top drug; (3) intra-class drug combinations are pairwise drug combinations within monotherapy (the same HF drug class); (4) inter-class drug combinations are pairwise drug combinations in combination therapy (between two different HF drug classes) and (5) with-comorbidity drug combinations are pairwise drug combinations that include a HF medication and a comorbidity medication. We consider HF patients with comorbid diabetes mellitus in this analysis.

Please refer to the Supplementary Note S7 for the data statistics of the drugs and combinations. Graph-Transformer generally performed better in predicting the drug response of combination therapy than single therapy (ρ = 0.01) (Figure 2D). Among the most frequently taken medications per drug class, Graph-Transformer performed the best on Atorvastatin from Statin (RMSE 0.0045), followed by Losartan from ARB (RMSE 0.0046) (Figure 3A). Graph-Transformer’s performance across all the intra-class drug combinations are comparable (ρ > 0.05) in Figure 3B. The inter-class drug combination results in Figure 3C suggest that Graph-Transformer predicts drug response statistically better for Lisinopril from ACEI in combination with medications from the other 4 drug classes. Graph-Transformer performed comparably (ρ > 0.05) in the with-comorbidity drug combination analysis Figure 3D.

Figure 3.

Drug class and combination evaluations. (A) Top drug per class, (B) inter-class drug combinations, (C) intra-class drug combinations, and (D) with-comorbidity drug combinations. Each box plot shows the distribution of the model’s performance on the 10-fold data associated with the respective cohort. The average RMSE score is denoted by the red triangle, the median score by the red horizontal line, and the corresponding P-value is annotated above each box plot. The degree of statistical significance is denoted by the number of asterisks “*” while not significant with “ns.”

Phenogroup characteristics and outcomes

Please refer to Supplementary Note S8 for details regarding the statistical analyses on the patient characteristics and outcomes. The four phenogroups were significantly different from each other both in terms of the patient characteristics (highlighted in red in Figure 4B y-axis) and outcomes. We compartmentalized these differences across the phenogroups to seven clinical parameters—BMI, Biomarkers, Hemodynamics, Comorbidities, Treatment 1, Treatment 2 and Outcome—for ease of comparison. BMI corresponds to the mean BMI value aggregated over the patients in a phenogroup. The Biomarkers includes BNP and NT-proBNP and Hemodynamics consists of the DBP, SBP and SVR variables, each represented with its mean value. The Comorbidities include 6 comorbidities (Atrial Fibrillation, Chronic Kidney Disease, Shortness of Breath, Hyperlipidemia, Diabetes Mellitus and Hypertension) and their prevalences. Treatment 1 corresponds to the percentages of patients taking medications under ACEI and BB within a phenogroup. Treatment 2 corresponds to the percentages of patients taking medications under ARB, Statin and LD. We categorized the relative level of each of these six parameters as either low or high depending on the respective normalized mean value/percentage. The Outcome is the median survival times of the phenogroup for the all-cause mortality in the Kaplan-Meier analyses and is categorized as low, high or mixed. Mixed indicates that the median survival times across the different treatments did not follow a general low or high trend for the phenogroup. Note that the categorizations of the parameters for each phenogroup are relative to the other phenogroups.

Figure 4.

Phenogroups generated with the phenotypical patient representations. (A) K-Means clusteringbased patient stratification. Each dot corresponds to a patient in the HF cohort and the color indicates the phenogroup it is assigned to. (B) Heatmap visualization of descriptive characteristics of patients in phenogroups. Data statistics have been normalized, so darker shade indicates a higher value relative to the lighter shade. The statistically significant (ρ < 0.05) characteristics are highlighted in red. (C) Kaplan-Meier curves to determine the association between phenogroups and all-cause mortality outcome for five HF treatments. Median survival times are denoted by the dashed lines.

The categorization of patients into distinct phenogroups based on these seven clinical parameters offers valuable insights for personalized treatment strategies. Each phenogroup reveals unique clinical challenges and opportunities that could guide targeted interventions. Phenogroup 1 is characterized by high BMI, low Biomarker levels, a high prevalence of Comorbidities, including the highest obesity rates (47.75%), and a high Outcome. Management strategies should prioritize weight control and cardiovascular therapies better suited to their metabolic profile. A favorable response to Treatment 2 (ARB, Statin, LD) suggests that lipid-lowering and blood pressure management interventions are particularly effective. Phenogroup 2 presents with low BMI and biomarker levels but high Hemodynamics, which might misleadingly suggest robust cardiovascular health. Instead, these characteristics likely indicate an inadequate response to treatment (low Outcome), reflecting uncontrolled hypertension with high DBP and SBP. This interpretation necessitates a reassessment of therapeutic strategies, focusing on enhancing treatment adherence and possibly exploring more potent or alternative therapies to improve outcomes. Phenogroup 3, with its high Biomarker levels and the highest rate of atrial fibrillation, indicates a high-risk cardiovascular profile. Despite low Hemodynamics and the lowest obesity prevalence (20.11%), traditional risk factors like weight may not predict outcomes effectively in this group. The low treatment response to both 1 and 2, in general, suggests that these patients derive less benefit from standard heart failure management. Prioritizing rhythm control may benefit this subgroup and warrants exploration. Phenogroup 4 exhibits high values across most parameters, indicating a complex patient profile with significant management challenges. High rates of diabetes and chronic kidney disease, coupled with high Biomarkers and BMI, call for a multifaceted approach to treatment that addresses both metabolic and cardiovascular needs. The favorable response to both Treatments 1 and 2 (high Outcome) suggests that these patients might benefit from aggressive multifactorial therapy to manage their chronic conditions and improve long-term outcomes. Two of our phenogroups are in part consistent with phenogroups identified in previous studies,^22,23 which include (1) Phenogroup 3 characterized by a high prevalence of atrial fibrillation but low prevalences of diabetes mellitus and obesity and (2) Phenogroup 4 characterized by high prevalences of obesity and diabetes mellitus.

Individual phenogroup performance

We evaluated the drug response prediction performances in RMSE among the four individual phenogroups across the 5 treatment classes using our graph-based model and compared them against the results obtained for the HF subtypes HFrEF and HFpEF. The comparative results were visualized as a heatmap in Figure 5, with better performance (low RMSE) shown in darker shades and statistical significance denoted by asterisk. Graph-Transformer generally demonstrated better drug response generalization for the four phenogroups than the HF subtype-based subgroups, with statistically significant RMSE reductions in ACEI, BB, Statin and LD evaluations.

Figure 5.

Comparative evaluation of the drug response prediction performance in RMSE between the existing HF subtypes (HFrEF and HFpEF) and the clustering-derived four phenogroups. The rows of the heatmap matrix correspond to the HF treatment the performance is evaluated for and the columns are the particular patient subgroup as the input to the model.

Discussion

In this study, we proposed a graph-based framework to predict treatment outcomes for heart failure patients by leveraging a large EHR database from Mayo Clinic. In particular, we were able to demonstrate that adapting the Transformer’s Self-Attention mechanism to the GNN model facilitates in learning phenotypical patient representations that can be exploited for patient stratification. The central concept of our approach is to simultaneously capture both the temporal dynamics and diverse dependencies present in EHR data. This is done to effectively track the evolving physiological trends and interactions throughout the progression of heart failure. By learning patient embeddings through this method, we can effectively stratify heart failure cohorts, leading to more precise predictive outcomes.

The results of this study demonstrate that the proposed graph-based model was able to outperform traditional drug response models by a significant margin (RMSE reduction of 79%, ρ = 0.001) due to the inability of those models to capture the complex, irregular topology of the EHR data. Empirical findings point out that the retrospective analysis of the clinical features in EHR, especially the lab tests, is informative in the accurate prediction of drug response. The lab tests DBP, SBP, and SVR are routinely collected and are considered modifiable risk factors imperative in HF prognosis,²⁶ indicating that the hemodynamic effects of the pharmacological intervention over time provide informative patterns for the model’s generalization. The assessment of single therapy revealed the best performance for the ACEI cohort, while combination therapy underpinned better predictive power for BB, ARB, Statin and LD. ACEIs have shown to be effective in the treatment of HF as they are generally recommended as the first-line therapy.²⁷ The role of NT-proBNP as a promising biomarker for assessing drug efficacy and its prognostic changes induced by ACEI medications are supported in multiple reports.^28,29

A novel approach in our research is the use of unique phenogroups to enhance the performance of drug response predictions. Our model significantly outperformed both the general patient spectrum and traditional clinical subtypes (such as HFrEF and HFpEF) (best mean RMSE of 0.0032). These findings support the feasibility of graph-based AI informed by EHR phenotypes to tackle the heterogeneity innate in HF’s pathophysiology. These phenogroups displayed phenotypically distinct profiles across 7 groups of clinical parameters (BMI, Biomarker, Hemodynamic, Comorbidities, Treatment 1, Treatment 2 and Outcome), suggesting different underlying mechanisms that could be targeted for effective intervention. Across all four phenogroups, we observed the following key findings: (1) the parameter Treatment 2 generally displayed a direct proportional relationship with Outcome. In other words, the higher usage of ARB, Statin or LD treatment within a phenogroup increased the survival of those patients. For example, Phenogroup 1 had a higher percentage of ARB usage (40.35%) than Phenogroup 3 (30.99%), and also a statistically higher (ρ = 4.6e-06) median survival time (1.17 years or 14 months) than the latter (0.64 years or 7 months). (2) Treatment 2 had a direct proportional relationship with BMI (generally) and Outcome. For example, Phenogroup 3 had a lower BMI (30.13) than Phenogroup 4 (33.52), a lower ARB usage (30.99%) than Phenogroup 4 (42.48%), and also a statistically lower (ρ = 3.3e-07) median survival time (0.64 years or 8 months) than the latter (1.29 years or 15 months). This finding was further corroborated through covariate analysis in Supplementary Note S9. Although higher BMI is a well-established risk factor for HF and is also associated with increased incidence of coronary artery disease, previous meta-analyses have shown it to be linked to improved survival in HF patients, generally termed as “obesity paradox”.³⁰ Additionally, randomized studies may be required to confirm our results and reconcile the conflicting findings in the literature. (3) Comorbidities had a direct proportional relationship with Treatment 2, and in turn with Outcome based on finding (1). This may be related to the fact that the occurrence of a comorbid condition with HF could lead to combination therapy. For example, ARB is administered as an antihypertensive to HF patients with hypertension.³¹ (4) Across phenogroups, Hemodynamics did not display a definitive relationship with Treatment 2 and thus had mixed Outcome. For example, Phenogroup 4 had a higher level of DBP or SBP that resulted in increased usage of Treatment 2 and improved survival. However, this relationship was inverse for Phenogroups 1 and 2. (5) Biomarker in general had a direct proportional relationship with Treatment 1, with a similar change in Outcome in general. For example, Phenogroup 4 had a higher NTproBNP (5432.13) than Phenogroup 2 (4715.60), and also a higher usage of BB (92.85%) than Phenogroup 2 (86.59%). Subsequently, the survival of Phenogroup 4 was higher (4.04 years or 46 months) than Phenogroup 2 (3.41 years or 40 months) (ρ = 0.05). (6) Comparisons of the patient survival between the treatments showed the highest survival with ACEI and lowest survival with ARB. This is the overall survival order ACEI > Statin > LD > BB > ARB, with mean survival times of 6.51, 5.69, 3.83, 3.10, 0.99 years, respectively, across all four phenogroups.

The unique characteristics of these phenogroups provide therapeutic insights toward tailored treatment strategies. ARB and Statins have shown significant benefits for older patients,^32,33 consistent with the favorable response to Treatment 2 observed in Phenogroup 1 (mean age of 74.78 ± 12.21; mode age 86) and Phenogroup 4 (mean age 74.42 ± 12.27; mode age 81). Conversely, some recent studies have found that lipid-lowering drugs might have adverse effects on older patients with heart failure.³⁴ However, our data indicate that Phenogroup 1, representing the oldest patients, showed positive responses to these treatments. This is supported by a recent study indicating that LD administration could improve 30-day outcomes in hospitalized patients.³⁵ Since ACEI and ARB work through the same pathways, the treatment responses to these two medications are generally expected to be comparable. Our results confirm this assumption, as Phenogroups 1, 2 and 4 exhibited similar responses to both Treatment 1 and Treatment 2. These similarities are consistent with clinical guidelines recommending ACEI or ARB for patients with kidney disease or diabetes,^36–38 which is reflected in the strong response from Phenogroup 4. In addition to these findings, a curious observation emerged in Phenogroup 2, which showed low responses to both Treatments 1 and 2, despite having the highest hemodynamics. This may indicate the need for reassessment of therapeutic strategies, focusing on improved treatment adherence and possibly exploring more potent or alternative therapies. Lastly, the poor response observed in Phenogroup 3, which has a high prevalence of atrial fibrillation, aligns with research that underscores the complex relationship between atrial fibrillation and heart failure.³⁹ Given the intertwining nature of these conditions, restoring sinus rhythm should be a primary goal.

Gender-related differences in the effectiveness of HF treatments have been reported in previous works.^40,41  Supplementary Figure S13 presents the treatment distributions based on gender across the four phenogroups in our dataset. The results suggest that a significantly higher number of male HF patients receive ACEI, BB, Statin and LD treatments than their female counterparts. ARB intake, however, was significantly more among female patients in Phenogroups 1 and 4. As ARB is typically taken as an antihypertensive among HF patients, our analysis further revealed (Supplementary Figure S14) that the prevalence of hypertension is more in females than males in Phenogroups 1 and 4. However, the differential treatment distributions between the gender do not necessarily correlate to better treatment effectiveness as findings from our covariate analysis (Supplementary Figures S8-S12) revealed statistically significant higher survival in females than males in relation to BB, ARB, LD and Statin treatments for Phenogroup 4.

The proposed Graph-Transformer model can predict outcomes for new patients who meet just one criterion—having longitudinal EHR data from at least two visits—without the need for phenotype grouping. As demonstrated in Figure 2B, this approach achieved better performance compared to existing methods. Additionally, if a patient exhibits the same phenotypic characteristics as those identified in our proposed phenogroups shown in Figure 4B, outcomes can be predicted using a more customized model tailored for these groups. To test the automated identification of new patients into our stratified groups, we employed two different approaches. In the first approach, we used K-Means clustering to predict phenogroup membership as it is a centroid-based clustering algorithm. We evaluated this on our patient graph representations dataset via 5-fold cross validation, where in each fold a training set is presented to K-Means method to first identify four clusters; then for each instance in the test set, the phenogroup is determined based on the nearest cluster centroid. Based on this approach, we were able to obtain a mean F1 score of 0.60 averaged across all folds. In the second approach, we optimized a machine learning model for phenotyping of new patients based on the clustering results. To demonstrate this on our data, we created labeled training and test sets with 80/20% split using the learned graph-based patient representations as the inputs and the K-Means predicted cluster as the corresponding labels. Random Forest classifier was subsequently trained on the training portion through 10-fold cross validation and its performance evaluated on the test set using F1. The classifier was able to accurately assign new patients to phenogroups 1, 2, 3 and 4 with F1 scores of 0.91, 0.83, 0.81 and 0.94, respectively. These results increase our confidence that the detected phenogroups will hold up in populations similar to the Mayo Clinic data used in this study. However, it is important to note that our model cannot predict outcomes for new patients who have not been previously seen and have not received any HF treatment.

In summary, the main strengths of this study include (1) we presented a data-driven computational model that not only learns from the patient records to accurately predict diverse treatment effects, but also produces phenotypical patient representations that can be stratified into unique phenogroups to help enhance the model’s generalization capability for more personalized pharmacological therapy. Traditional computational phenotyping methods typically identify subtypes by analyzing phenotypical feature embeddings in the context of diagnosis or prognosis.^23,42 However, the validity of these identified or discovered subtypes often remains unconfirmed, rendering the outcomes of these computational methods less applicable in guiding clinical practices. Our study diverges from this by not just discovering subgroups but demonstrating their utility in patient stratification for enhanced computational prediction. (2) We used a sufficiently large HF cohort (11 627 patients) to develop and evaluate our drug response model and further validated the phenomapping results on two validation cohorts to ensure robustness of the derived phenogroups, as presented in Supplementary Note S10. We have demonstrated that our model can be applied to new patients with the same variables in the Mayo Clinic environment. However, we recognize that the phenogroups require further external validation to confirm their generalizability across broader settings. (3) We modeled the patient’s EHR history as graph input using the Transformer and GNN models end-to-end, that more accurately captures the clinical events and their temporal-heterogeneity interactions in the patient’s HF trajectory. (4) We evaluated the efficacy and viability of the proposed graph-based framework for drug response prediction through extensive quantitative and phenomapping experiments across a wide range of HF treatments. Our key findings elucidate insights about associations among different clinical parameters across the phenogroups; this could help improve understanding of the phenotypic heterogeneity of HF in relation to the treatment outcome for better patient care and a potential for more effective allocation of healthcare resources in managing heart failure.

The study includes some limitations that warrant further investigations. First, this is a retrospective study where 94% of our patients are Non-Hispanic White. This limits validity and applicability of the model to more diverse HF populations. Second, All patient records utilized in this study were sourced from the Mayo Clinic. This presents a challenge in terms of the model's generalizability as different institutions often have customized EHR systems, each with its unique set of patient features and data distributions. These differences can lead to domain shifts, which might affect the model's performance when applied to data from other institutions. To truly harness the potential of our framework and ensure its applicability in a real-world clinical setting, it's imperative to adapt it for cross-institutional predictions. By incorporating data from a more diverse population, such as what's available in projects like the All-of-US,⁴³ we can gain a more comprehensive understanding of HF's heterogeneity. Future research using multiple EHR datasets, especially prospective studies, may help further corroborate and expand the findings in our current study.

Conclusion

This work successfully demonstrated the feasibility of graph-based AI with EHR data for treatment outcome analysis in HF patients. Importantly, application of unsupervised clustering approach facilitated patient stratification into distinct phenogroups with differential clinical characteristics and outcomes. This study focused on achieving predictive improvement using the phenogroups, that underscores their potential to enhance personalized prognostic predictions.

Author contributions

S.C. and N.Z. jointly conceived the project. S.C. conducted the data collection and processing, with input and consultation from Y.C., Q.D., and P.L. The methodology, including computer coding, experimentation and results analysis, was developed by S.C. All authors contributed with experimental feedback and participated in editing the manuscript. The final version of the manuscript was approved by all authors. N.Z. provided overall supervision for the project.

Supplementary material

Supplementary material is available at Journal of the American Medical Informatics Association online.

Funding

This study is supported by the National Institute of Health (NIH) NIGMS (R00GM135488).

Conflicts of interests

The authors declare no competing interests.

Data availability

Protected Health Information (PHI) restrictions apply to the availability of the clinical data here, which were used under IRB approval for use only in the current study. As a result, this dataset is not publicly available. Qualified researchers affiliated with the Mayo Clinic may apply for access to these data through the Mayo Clinic Institutional Review Board.

Code availability

The code used to train and generate results can be found at https://github.com/bioIKEA/HF_response_regression.