Preliminary insights into artificial intelligence guided dosing in hypertension and diabetes: challenges and lessons learnt in a pilot feasibility study

Jennifer Sumner, Mehul Motani, Jina Ali, S. H. Lee, Danliang Ho, Dylan Tan, Lieng Hsi Ling, Andre Teck Huat Tan, Gim Gee Teng, Santosh Kumar Seetharaman, Satya Pavan Kumar Gollamudi, Lin Siew Chong, Amartya MukhopadhyayJAMIA 2026 · DOI

TL;DR

This study explores the feasibility of using CURATE.AI, an AI platform for personalized drug dosing, in an outpatient setting for hypertension and diabetes. While demonstrating potential, the pilot highlights challenges in data acquisition for calibration, suggesting future work should focus on patient groups with greater dosing variability.

What Problem Does This Paper Address?

Precision medicine aims to tailor treatments to individual patient characteristics, but optimizing drug dosages, especially for chronic conditions like hypertension and type II diabetes, is complex. Existing AI approaches have shown promise in inpatient settings, particularly for chemotherapy, but their application in outpatient chronic disease management is less explored. Outpatient settings present unique challenges: less direct clinician control over patient actions, the lifestyle-driven nature of chronic diseases leading to variability, and difficulties in consistently collecting clinical response data outside of scheduled visits. This paper investigates whether an AI platform, CURATE.AI, can be practically implemented in this setting to guide personalized drug dosing, addressing the need for more efficient and precise chronic disease management.

Methods

The researchers conducted a single-arm feasibility study using the CURATE.AI platform. CURATE.AI works by creating an individualized profile based on a patient's drug dose-response relationship, using a quadratic equation to model treatment response. To calibrate the system for each patient, a minimum of three data points linking drug dosage to clinical response (blood pressure or blood glucose) were required.

The study involved two phases:

  1. Calibration Phase: This phase focused on gathering the necessary data to build the patient's CURATE.AI profile. This involved reviewing past medical records and scheduling up to four clinic visits for data collection. Home monitoring of blood pressure and continuous glucose monitoring were also employed. If sufficient calibration data couldn't be obtained, the participant exited the study.
  2. Active Phase: Participants who successfully calibrated entered this phase, where CURATE.AI generated monthly dosing recommendations. Physicians then reviewed these recommendations and made the final dosing decisions. The study tracked the alignment between physician and AI recommendations.

Data analysis involved evaluating feasibility criteria such as recruitment and completion rates, adherence to clinic visits and home monitoring, patient satisfaction, consistency of CURATE.AI recommendations across different data sources, and the alignment of AI recommendations with physician decisions.

Datasets

The study utilized clinical data from patients recruited from Alexandra Hospital's outpatient clinics. This data included:

  • Hypertension Data:
    • Single systolic pressure readings
    • Single Mean Arterial Pressure (MAP) readings
    • Mixed data combining home monitoring and single clinic readings for systolic pressure and MAP.
  • Diabetes Data:
    • Glycated hemoglobin (HbA1c) readings
    • Continuous Glucose Monitoring (CGM) data (including time in range and glycaemic variability)

No new benchmarks were introduced; the study focused on evaluating an existing AI platform with real-world patient data.

Key Insights and Ideas

  • Feasibility in Outpatient Chronic Care: The study successfully demonstrated that CURATE.AI can be deployed in an outpatient setting for chronic disease management, indicating a potential pathway for AI-driven personalized dosing beyond acute care.
  • Calibration Data is King (and Hard to Get): The primary hurdle was obtaining sufficient, high-quality calibration data. Patients with stable chronic conditions often require minimal dose adjustments, limiting the data points needed for effective AI model training. This suggests that AI dosing tools might be more immediately applicable to patients with greater initial dosing variability.
  • Physician Buy-in and Workflow Integration: While physicians were involved in the decision-making process, the AI interface was not directly accessible to them, requiring a research team intermediary. This highlights the importance of user-friendly interfaces and seamless integration into existing clinical workflows for broader adoption.
  • Data Source Variability Matters: Different data sources (e.g., single clinic readings vs. home monitoring, different analytical approaches) led to varying CURATE.AI recommendations, creating uncertainty about which data input is most reliable. This underscores the need for further research into optimizing data collection and analysis strategies.
  • Patient Satisfaction is High: Despite adherence challenges, participants reported high satisfaction with their monitoring and care, suggesting that patients are receptive to technology-assisted health management.

Key Results

  • Recruitment: Eighteen participants consented, achieving 45% of the target recruitment.
  • Study Completion: Thirteen participants completed the study.
  • Active Phase Conversion: Only three participants progressed to the active phase, primarily due to insufficient dose adjustments during calibration.
  • Adherence: Adherence to scheduled clinic visits was 76%, and home monitoring adherence averaged 81%. Barriers included technical issues, work conflicts, and perceived lack of necessity for further appointments due to clinical stability.
  • Patient Satisfaction: Participants reported high satisfaction (≥88%) with monitoring and care.
  • Dosing Recommendations: Actionable dosing recommendations were generated for two of the three participants who reached the active phase. The alignment with physician decisions varied depending on the data source used. For one hypertensive patient, AI recommendations consistently matched physician decisions. For a diabetic patient, recommendations differed in 3 out of 4 follow-up visits.

Limitations and Future Work

Limitations: * Small Sample Size: The study had a limited sample size (n=17 for analysis), which restricts the generalizability of the findings. * Physician Interface: Physicians did not have direct interaction with the CURATE.AI tool, as the interface was not sufficiently developed for deployment. This limits insights into the direct user experience. * Low Active Phase Conversion: Only three participants reached the active phase, meaning the findings from this phase are based on a very small number of cases. * Data Source Uncertainty: The study identified variability in recommendations based on different data inputs, highlighting a need for optimization. * Split Dosing Regimens: Handling AM/PM split dosing regimens presented practical challenges for calibration.

Future Work: * Target Newly Diagnosed Patients: Future studies should focus on newly diagnosed patient groups who are more likely to experience greater dosing variability and require more frequent titration, thus facilitating better AI model calibration. * Optimize Data Requirements: Further research is needed to align data requirements with patient and clinical characteristics, and to determine the most suitable data sources and formats for AI-guided dosing. * Explore Multiple Drug Optimization: Investigating the practicalities of concurrently optimizing multiple medications with CURATE.AI is suggested. * Develop User-Friendly Interfaces: Creating direct interfaces for clinicians to interact with AI dosing tools is crucial for adoption. * Address Tablet Dose Limitations: Exploring solutions like 3D drug printing could offer more precise dosing increments, overcoming limitations of current tablet formulations.

Summary updated

Abstract

This pilot demonstrates the feasibility of deploying CURATE.AI into outpatient care but underscores the importance of aligning data requirements with patient and clinical characteristics. Future studies should target newly diagnosed patient groups with greater dosing variability to optimise calibration and assess clinical utility.
BibTeX
@inproceedings{sumner2026preliminary,
  author = {Jennifer Sumner and Mehul Motani and Jina Ali and S. H. Lee and Danliang Ho and Dylan Tan and Lieng Hsi Ling and Andre Teck Huat Tan and Gim Gee Teng and Santosh Kumar Seetharaman and Satya Pavan Kumar Gollamudi and Lin Siew Chong and D. Heffernan Ho and Amartya Mukhopadhyay},
  title = {{Preliminary insights into artificial intelligence guided dosing in hypertension and diabetes: challenges and lessons learnt in a pilot feasibility study}},
  booktitle = {JAMIA},
  year = {2026},
  doi = {10.1093/jamiaopen/ooaf153},
}
Figure