Disease Risk Prediction Using Structured EHR Data: Can Generalist Large Language Models Match Specialized Clinical Foundation Models? A Comparative Evaluation with Fine-Tuning
1McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston, Houston, TX, USA
2Department of Biomedical Informatics and Data Science, Yale University, New Haven, CT, USA
*Corresponding author laila.rasmy.gindybekhet@uth.tmc.eduAbstract
Background
Electronic health records (EHRs) with clinical decision support tools are now ubiquitous in healthcare organizations. Clinical foundation models (CFMs) pretrained on large-scale, heterogeneous structured EHR data have emerged as a powerful approach to improve predictive performance and generalizability. Meanwhile, large language models (LLMs) pretrained on broad data sources are being applied to an expanding range of healthcare tasks. However, it remains unclear whether generalist LLMs can match specialized CFMs for disease risk prediction using structured clinical data.
Methods
We compared CFMs (Med-BERT, CLMBR) against fine-tuned generalist LLMs (Mistral, LLaMA-2/3/3.1), a clinical LLM (Me-LLaMA), and LLM-generated embeddings paired with simple classifiers (using DeepSeek, Qwen3, and GPT-OSS) on two disease risk prediction tasks: heart failure risk among diabetic patients (DHF) and pancreatic cancer diagnosis (PaCa). Evaluations spanned multi-site EHR data, claims data, and an open-source single-institution benchmark (EHRSHOT). Performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC).
Results
On larger EHR and claims cohorts (>30,000 patients), fine-tuned CFMs outperformed fine-tuned LLMs by a small but statistically significant margin (<1% AUROC). The clinical LLM performed comparably to generalist LLMs despite being smaller. On the open-source PaCa cohort (3,810 patients, 199 cases), LLMs achieved slightly higher AUROCs that were not statistically significant (LLaMA-3.1-70B 86.1% vs. Med-BERT 85.3%, p=0.27), but CFMs achieved significantly higher AUPRC (Med-BERT 55.9% vs. LLaMA-3.1-70B 41.1%, p=0.001). Notably, LLM-generated trajectory embeddings paired with logistic regression or a simple MLP, without any LLM fine-tuning, achieved the best overall performance, with AUROC exceeding 90% (Qwen3) and AUPRC reaching 66% (GPT-OSS 20B).
Conclusion
LLM-generated embeddings with lightweight classifiers outperformed both fine-tuned CFMs and fine-tuned LLMs on AUROC and AUPRC. While these results demonstrate the potential of generalist models to match or surpass specialized CFMs, their substantially greater computational cost and variable AUPRC performance in the fine-tuning setting warrant caution. We provide a reproducible evaluation framework and codebase to support continued benchmarking.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Funding Statement
This study received funding in part from the National Institutes of Health and the National Library of Medicine grant R01LM014249.
1.Introduction
Clinical decision support tools based on predictive models using structured electronic health records (EHRs) are commonly in use in healthcare organizations as they support early detection1, personalized treatment2, and resource optimization2. Over the past decade, machine learning-based predictive models have become the standard approach in clinical decision support. Pretrained models designed specifically for structured EHR data, such as Med-BERT3, BEHRT4, CLMBR5, TransformEHR6, and EHRMamba7, were developed mainly to improve the performance and the generalizability of such predictive models even with commonly known limitations of structured EHR data such as missingness, noise, inaccuracy, and sparsity. These models address these challenges primarily by generating contextual embeddings that represent the patient trajectory based on the parameters learned from complex patterns present in high-dimensional and longitudinal EHR data for millions of patients during pretraining.
Foundation models are defined as large AI models pretrained on broad data at scale to enable adaptability across a wide range of downstream tasks8. Following this terminology, we refer to pretrained models specifically designed for structured EHR data as clinical foundation models (CFMs), a class of specialized foundation models that have shown strong predictive performance on EHR-based tasks □. Meanwhile, generalist foundation models, especially large language models (LLMs), have disrupted many fields, including healthcare10–12, due to their adaptability and ability to generalize across diverse tasks. As awareness of these models’ capabilities grows, decision makers are increasingly exploring the use of LLMs for diagnostic decision-making10–12. This raises an intriguing question: Can LLMs outperform CFMs, boosting the performance of clinical predictive models using structured EHR data?
Evidence directly comparing LLMs to specialized CFMs remains limited. CFMs are explicitly designed and optimized for structured EHR data and have shown over 20% improvements in prediction accuracy3 and require less task-specific training data for fine-tuning when needed. On the other hand, generalist LLMs like GPT13,14 and LLaMA15,16, are pre-trained on various datasets and handle free-text inputs. Clinical LLMs (CLLMs) like Me-LLaMA17 are specialized versions of generalist LLMs, which are fine-tuned on domain-specific clinical text data, mainly clinical notes, also available in EHRs.
In the context of clinical prediction tasks, recent studies explored the potential of LLMs and compared their performance against specialized machine learning (ML) models. Chen et al10 and Hu et al.18 found that LLMs cannot beat traditional ML models in clinical prediction. CFMs have been shown to improve the performance of traditional ML models like logistic regression and deep learning models in many instances3,5,9,19,20. Traditional ML models may still perform well, particularly when trained on small local datasets, although they may tend to overfit. Shoham et al.11 and Acharya et al.12 fine-tuned LLaMA models on structured EHR data, reporting modest improvements over their respective versions of Med-BERT. Notably, these versions were pre-trained on a single source, such as MIMIC-IV, using relatively small cohorts compared to the scale typically required to pre-train an optimal CFM.
Unlike previous works, in this study we focused on comparing the utility of CFMs pre-trained on millions of patients’ full trajectories and LLMs for disease risk prediction, which commonly needs modeling for complex, long-term interactions within patient histories3,11,12. We restricted our evaluation to open-weight LLMs for two reasons: first, open-weight models allow supervised fine-tuning and embedding extraction on local infrastructure, which is essential for reproducibility; second, they avoid transmitting protected health information to external APIs, an important consideration for HIPAA compliance and institutional data governance in clinical settings. We compared the prediction discriminative accuracy of classifiers built on top of finetuned CFMs, such as Med-BERT and CLMBR5, versus fine-tuned LLMs such as Mistral21 and LLaMA15,16 and a CLLM Me-LLaMA17, for predicting patient risk of getting diagnosed with pancreatic cancer (PaCa) and the diabetic patients risk of developing heart failure (DHF) which using the same evaluation cohorts used in the Med-BERT3 paper as well as the PaCa cohort available in the EHRSHOT19 dataset. In this study, our contributions are summarized as follows:
- We present the first head-to-head comparison of open-weight LLMs against CFMs pretrained at scale—including an encoder-based CFM trained on over 50 million patients’ longitudinal EHR and claims data (Med-BERT v2) and a decoder-based CFM trained on 2.57 million patients’ EHRs (CLMBR)—for disease risk prediction.
- We evaluate the impact of different data serialization strategies for converting structured EHR data into textual input for LLMs.
- We compare two modes of leveraging open-weight LLMs—supervised fine-tuning versus embedding extraction with lightweight classifiers—and show that the latter achieves the strongest overall performance.
2.Methods
2.1Models
We evaluated three model categories: CFM, generalist LLM, and CLLM. CFMs are a specialized type of foundation model (FM) developed for electronic medical record (EMR) data, designed to address healthcare-specific tasks9. We tested two representative CFMs: Med-BERT and CLMBR. Med-BERT is an early encoder-based CFM trained on longitudinal structured EHR data from large-scale clinical datasets, and has been shown to substantially improve prediction accuracy for disease prediction tasks3. CLMBR is a decoder-transformer-based, autoregressive foundation model (CLMBR-T-base), originally described by Steinberg et al. (2021)5 and later improved by Wornow et al. (2023)9. For LLMs, five open-source generalist LLMs were selected for fine-tuning: the Mistral-7B-Instruct-v0.221 model and LLaMA15,16 series models, including LLaMA-2-13b-hf, LLaMA-3-8B-Instruct, LLaMA-3.1-8B-Instruct, and LLaMA-3.1-70B-Instruct. Mistral-7B incorporates grouped-query attention (GQA)22 for faster inference by clustering related queries and sliding-window attention (SWA)23 to efficiently handle long sequences. LLaMA-215 is an auto-regressive language model that features a standard Transformer architecture with extensions, including root mean square layer normalization (RMSNorm)24 for pre-normalization and a swish-gated linear unit (SwiGLU)25 activation function. It was fine-tuned using reinforcement learning from human feedback (RLHF)26. Finally, we included Me-LLaMA17, a CLLM built on LLaMA-2, which is continually pre-trained on 129B tokens and instruction-tuned with 214K clinical samples. Full model details and comparison of the key characteristics are provided in Appendices A and B of the supplementary materials. Additionally, as the LLM domain is fast-moving, we evaluated a group of the latest LLMs to directly generate patient representations, without fine-tuning, and trained a classifier on top of those pre-generated patient embeddings. For patient representations, we tested specialized LLMs such as MedGAMMA27 and Phi428, and generalized LLMs with reasoning capabilities such as Deepseek29, GPT-OSS30, Qwen331 thinking, and LLM embedding models such as Qwen3 8B.
2.2Datasets and prediction tasks
This study focuses on two binary classification disease prediction tasks: DHF and PaCa, mirroring those addressed by Med-BERT3. The objective of these tasks is to predict whether a patient will be diagnosed with DHF or PaCa based on their structured EHR history. The prediction relies on patient history, including past diagnoses, medications, procedures, and other EHR information. To ensure a fair comparison, we used the same datasets evaluated in the Med-BERT study. These include two datasets derived from a structured EHR database (DHF-EHR and PaCa-EHR), and a patient-level claims dataset, PaCa-Claims. We also included the publicly available EHRSHOT dataset for the PaCa prediction task to further expand the evaluation. To explore the impact of data richness, we processed patient histories in three levels of clinical events categories: (1) diagnosis-only (D) includes only diagnosis information, which is the same as Med-BERT; (2) the combination of diagnosis, medication, and procedure (DMP), a richer representation similar to the Med-BERT v232 preprocessing; and (3) all-codes (ALL), which incorporates all available structured data, such as demographics, vitals, laboratory results and clinical observations. Appendix C in the supplementary materials includes a detailed description of datasets and prediction tasks.
2.3Data preprocessing
Structured EHR data were preprocessed to align with each model’s requirements. For Med-BERT, we followed the original preprocessing pipeline3, organizing patient visits chronologically and standardizing input lengths. For CLMBR, we used EHRSHOT scripts with minor modifications to ensure one prediction per patient and consistent prediction timing. LLM inputs were generated by converting structured data into text, following formats from CPLLM11 and LLaMA2-EHR12. We tested two prompt styles, the LLaMA2-EHR12 format with diagnostic frequency, to assess the impact on performance. For LLM embedding generation, due to run-time constraints, we generate embeddings for the first 4096 tokens of a patient’s history, including all visits and codes associated with each visit, ordered from most recent visit to oldest. Full preprocessing details are in Appendix D.
2.4Fine-tuning and evaluation
All models were fine-tuned using the HuggingFace Transformers library. Med-BERT and CLMBR were pretrained on GPU clusters, with fine-tuning following a similar setup but with task-specific adjustments. LLMs and the CLLM were fine-tuned using parameter-efficient techniques, including low-rank adaptation of large language models (LoRA)33 from the parameter-efficient fine-tuning (PEFT)34 techniques and model quantization to reduce memory usage and training time.
Following the Med-BERT paper and other previous works on clinical prediction tasks11,12, we used the area under the receiver operating characteristic curve (AUROC) as the main evaluation metric for all experiments. For the DHF-EHR, PaCa-EHR, and PaCa-Claims tasks, which are the same as those in Med-BERT, we used diagnosis information as input, evaluated three traditional machine learning models: logistic regression (LR), random forest (RF), and light gradient boosting machine (LGBM). All five generalist LLMs and the CLLM were fine-tuned and evaluated on these three tasks. For the PaCa task on EHRSHOT, we evaluated three combinations of clinical events categories and included CLMBR, Med-BERT v2, the CLLM, and the two best-performing LLMs from the DHF-EHR, PaCa-EHR, and PaCa-Claims experiments. Full fine-tuning and evaluation details are provided in Appendices E and F, respectively.
2.6Framework overview
Figure 1 shows the overview of the framework for each disease prediction task, as well as a summary of the results from selected experiments. Note that the patient information shown in the figure is not real patient data but just for illustration purposes. The process begins with the extraction of structured EHR data, which comprises both numerical medical codes and their corresponding textual descriptions. CFMs reformat the numerical codes following the data preprocessing methods of Med-BERT or CLMBR, while LLMs process the textual descriptions using prompt designs similar to CPLLM or LLaMA2-EHR. After preprocessing, both CFMs and LLMs are fine-tuned to generate predicted risk scores, which are subsequently evaluated using AUROC and AUPRC. Based on our findings, we also directly generate patient-level embeddings using different LLM models and file formats.
A detailed code repository to reproduce our experiments can be found at: https://github.com/ClinicalFM/LLM2Predict.
3.Results
3.1Descriptive analysis
Table 1 summarizes the descriptive analysis of the datasets and tasks included in this study, highlighting differences in patient sample size, visit frequency, and the scope of information used for modeling. Detailed descriptive analysis can be found in Appendix G of the supplementary materials.
3.2Fine-tuning results
Table 2 summarizes the average AUROC and standard deviations for different models across the DHF-EHR, PaCa-EHR, and PaCa-Claims tasks using the LLaMA2-EHR prompt format, offering important insights into model performance. Table S3 in Appendix H of the supplementary materials shows the model performance of tasks using the CPLLM prompt format, which consistently provides lower AUROCs than the LLaMA2-EHR format. The ML and CFM results in Table 2 are adapted from the results table of the Med-BERT paper3. Med-BERT + Bi-GRU achieved the highest AUROC (85.39) on DHF-EHR, while LLaMA-3.1-70B-Instruct and Me-LLaMA performed comparably (84.73 and 84.46). For PaCa-EHR, LLaMA-3.1-70B-Instruct and Me-LLaMA led with AUROCs of 82.96, slightly outperforming Med-BERT. Med-BERT remained best (80.57) in the PaCa-Claims task. P-values confirmed CFMs matched or outperformed LLMs significantly.
Table 3 explores model performance for the PaCa-EHRSHOT task under three levels of clinical event categories: D (only include the diagnosis information), DMP (include diagnosis, medication, and procedure information), and ALL (include all available information). Since the original Med-BERT3 model was pre-trained only using the diagnosis information, we selected to test Med-BERT v232, which was pre-trained on diagnosis, medication, and procedure information. Additionally, we only tested the better prompt format and the best generalist LLM based on the previous experiments, which are the LLaMA2-EHR prompt format and the LLaMA-3.1 series models. The results indicate that the LLaMA-3.1-70B-Instruct achieved the highest AUROCs in D (86.1) and DMP (86.65), while LGBM led in ALL (90.26). CFMs like Med-BERT v2 and CLMBR remained competitive to be the second best in D (85.25) and DMP (86.33). AUPRC results indicated that CFMs often outperformed LLMs, especially in D and DMP categories.
Moreover, Tables S4 and S5 in Appendix I of the supplementary materials compare the fine-tuning process speed by hours for all generalist and clinical LLMs. Mistral-7B was fastest among LLMs (42 hours), while LLaMA-3.1-70B required up to 190 hours. In contrast, CFMs completed fine-tuning in under 1 hour. The results summary in Figure 1 shows the results from DHF-EHR and PaCa-EHRSHOT (D), including the comparisons of AUROC, fine-tuning time, and number of trainable parameters for different models. The yellow area at the upper left corner of each subplot represents the models given the highest AUROC with the lowest fine-tuning time.
Table 4 shows that we achieve the best pancreatic cancer prediction performance using LLMs as encoders, without the need for fine-tuning. We tested various machine learning and basic MLP prediction heads and found that a simple logistic regression (AUROC 79.9% - 89.9%) or MLP (78.8% - 87.1%) yielded the best performance compared to tree-based (66.9% - 82.9%). We also found that adding instructions was commonly associated with better performance (67.7% - 89.9%) compared to without instructions (66.9% - 88.5%), and that using Markdown with a patient summary at the top was more effective (70.3% - 89.9%) compared to using the patient trajectory expanded in XML format (66.9% - 85.2%) (Supplementary Table 7). Additionally, we found that the latest embedding models, such as Qwen3, even with a lower number of parameters (8B) (85.1% - 90.8%), showed better performance than other instruction and thinking variant models with slightly larger parameters (20B) (82.8% - 88.4%). These results demonstrate that newer open-source models can efficiently generate helpful patient representations that, with a simple logistic regression model on top of it, get strong PaCa prediction results.
4.Discussion
In this study, we presented a comparative evaluation of generalist LLMs and specialized CFMs for disease risk prediction using structured EHR data from different real-world sources. We focused on two disease risk prediction tasks, Heart Failure for diabetic patients and Pancreatic Cancer, as our use cases, across three datasets from different sources: (1) EHR data from multiple health systems across the US (DHF–EHR, PaCa-EHR), (2) claims data (PaCa-Claims), and (3) an open-source cohort (PaCa-EHRSHOT), to support reproducibility and future benchmarking. We also tested three different combinations of clinical event categories on EHRSHOT, i.e., diagnosis codes only (D), diagnosis, medications, and procedures combination (DMP), and all codes available in the EHRSHOT dataset (ALL), including demographics, vitals, laboratory results, and clinical observations in addition to the previously mentioned clinical event types. While LLMs show promising results, CFMs and, sometimes, even simple ML models achieve slightly better discriminative accuracy, especially when fine-tuned on larger cohorts. For example, Med-BERT reached an AUROC of 85.4 for DHF prediction, outperforming LLaMA-3.1-70B (AUROC=84.7). Additionally, CFMs achieved significantly higher AUPRCs than LLMs in all EHRSHOT experiments, even when LLMs showed slightly better AUROC. However, when utilizing LLMs to generate patient-level embeddings for a text summary of their full trajectory, they lead to the best performance (for example, PaCa-EHRSHOT-ALL: AUROC of 90.8 and AUPRC of 64.7 using a 3-layer MLP on top of Qwen3 embeddings). Compared to previous studies that mostly compare LLMs to ML models10–12,18,35, our work contributes to benchmarking against specialized CFMs trained on tens of millions of patients’ data, such as Med-BERT, for disease risk prediction, while also highlighting how data representation, patient information formats, and LLM utilization modes influence model performance.
From our experiments, we observed several key findings. First, converting patient trajectories into text using the LLaMA2-EHR12 format led to approximately 1% higher discriminative accuracy compared to the CPLLM11 format (Supplementary Appendix H) and accordingl,y we followed the LLaMA2-EHR 12 format for all experiments reported in Tables 2 and 3. This improvement may be partly due to the inclusion of additional information such as event counts. We also noted that differences in token granularity, e.g. SentencePiece15 used in LLaMA-2 versus Tiktoken16 in LLaMA-3, are among the factors impacting the variability in performance as well as computational resource needs. Second, we found that in the majority of our experiments except for PaCa-Claims, LLaMA-3.1-70B is achieving the best results among all other LLMs tested, given it was the largest model we fine-tuned in this study. However, the difference was just marginal (<0.3%). Third, we found that the CLLM (Me-LLaMA), which was further trained on clinical text, performed comparably to generalist LLMs. For example, Me-LLaMA is leading to around 83% AUROC for the PaCa-EHR experiment which is the same as the LLaMA-3.1-70B model performance, even though it is a smaller model (13B). Med-BERT achieved the highest AUPRC (55.9%) on the PaCa-EHRSHOT (D) task, outperforming CLMBR (DMP) (54.9%) and LGBM (ALL) (52.0%). In contrast, the best-performing LLM (LLaMA3.1-70B on DMP) reached only 49.1%. This may be partly attributed to the quality of terminology mapping: medication codes mapping from RxNorm to Multum identifiers is nearly complete (90%), followed by mapping the diagnosis codes between standard OMOP SNOMED codes to ICD codes (82%), but the mapping of the procedure codes was just at 10%. Notably, Med-BERT v2 was trained only on diagnosis, medication, and procedure codes, and thus was excluded from the all-codes experiments (Table 3). Similarly, as we could not map a high percentage (>60%) of our multi-site EHR or claims data to the CLMBR vocabulary, we decided not to report the CLMBR model performance in Table 2. Lastly, while we only used LORA adaptors with the default configuration during the LLMs’ supervised fine-tuning (SFT) in this study’s experiments for computational efficiency, the SFT of the LLM models was at least 40x longer than the SFT of CFM and it required at least double the computational resources. For example, for PaCa-EHRSHOT which is our smallest cohort, the fine-tuning of Med-BERT and CLMBR took less than 5 minutes using a single GPU, while LLaMA2-3.1-8B (13M tunable parameters) took around 4 hours using 2 H100 80GB GPUs on the DMP experiment, Me-LLaMA (13B with 26M tunable parameters) took around 7 hours, and LLaMA2-3.1-70B (65M) took 18 hours. Nonetheless, the LLaMA2-3.1-70B (ALL) took 110 hours using 4 GPUs (Supplementary Appendix I).
While CFM is currently performing better on disease risk prediction tasks using longitudinal structured EHR data given the data sources, modalities, and pretraining tasks used to train such models, generalist LLMs have proven ability to outperform on many clinical tasks, such as question answering, clinical note summarization, automatic medical coding, and clinical trial matching36. Recent advances further demonstrate their potential: Health-LLM37 achieved strong multimodal health prediction by combining contextual and physiological data, reaching performance comparable to larger models like GPT-3.5 and GPT-4; prompt engineering and retrieval-augmented generation (RAG) can further enhance their performance on medical tasks38–40. Additionally, studies such as Acharya et al.12 and Beaulieu-Jones et al.35, showed that LLM models fine-tuned on clinical data can outperform traditional ML models for clinical prediction tasks. However, studies such as Chen et al. (2024)10 and Brown et al. (2025)41 have demonstrated the opposite. We are one of the early studies comparing LLMs to CFMs trained on millions of patients records including Med-BERT v2 trained on more than 50 million patients data, and our findings echo Hegselmann et al. (2025)42, showing that LLMs can achieve comparative performance to specialized models. They compared LLM-embedding models (GTE-Qwen2-7B-Instruct and LLM2Vec-Llama3.1-8B-Instruct) versus CLMBR on the 15 clinical prediction tasks from the EHRSHOT benchmark and UK Biobank, and found that LLM encoders can outperform CLMBR in disease onset prediction, although they did not specify the performance on PaCa cohort. Our findings demonstrate the promise of the use of LLMs for disease prediction. However, instead of completely replacing CFMs, future work should consider integrating those models to get the advantage from both. For example, LLM-based embeddings can be utilized while training CFMs’ to further enhance their performance and generalizability as proposed by Su et al. (2025)43.
However, this study is not without limitations. First, our evaluation is limited to binary classification-based disease prediction tasks, and it remains unclear whether the observed trends extend to multi-class or regression-based clinical predictions. Second, while we used multiple datasets, the scope of patient information was limited, and including a wider range of clinical features could potentially improve model performance. Third, the models evaluated in this study are based on structured EHR data, and their performance on unstructured data, such as clinical notes, remains unexplored. Future work could expand on this by evaluating the impact of context length on model performance, as explored by Wornow et al. (2025)44, or by investigating hybrid approaches that combine CFMs and LLMs, as proposed by Hegselmann et al. (2025)42. In conclusion, our study has shown that generalist LLMs have the promise to improve the performance of disease prediction tasks. However, there are a lot of aspects that need to be considered before fully relying on them, that includes how the data will be presented to the LLM, key parameters considered during the LLM fine-tuning, computational resources available, data privacy and HIPAA compliance especially for open-source models, and the main metrics crucial to evaluate the clinical utility of the fine-tuned model. While that highlights the idea that LLMs are an exciting new direction in clinical prediction, CFMs continue to play an important role in healthcare and should not be overlooked. Future research should continue to refine the application of LLMs in clinical settings and explore ways to integrate the strengths of both LLMs and CFMs for improved disease prediction.
Supporting information
Data Availability
All data produced in the present study are available upon reasonable request to the authors.
Institutional Review Board Statement
All data used in this study were subject to oversight and approval by the Committee for the Protection of Human Subjects (UTHSC-H IRB) under protocol HSC-SBMI-13-0549.
Data and Code Availability
For reproducibility, our codebase and the trained models will be available at https://github.com/ClinicalFM/LLM2Predict; however, for patient data privacy, confidentiality, and ethical reasons, the model training data cannot be shared. However, to reproduce results on the EHRshot cohort, access can be requested through https://som-shahlab.github.io/ehrshot-website/.
Funding Acknowledgments
This study received funding in part from the National Institutes of Health and the National Library of Medicine grant R01LM014249.
Conflict-of-Interest Statement
The authors declare no conflicts of interest.