Artificial intelligence guided dosing decisions: a qualitative study on health care provider perspectives

Jennifer Sumner, Jina Ali, Mehul Motani, Abigail Ang, Dean Ho, Amartya MukhopadhyayBMJ 2025 · DOI

TL;DR

While AI promises to move medicine past the "one-dose-fits-all" model, its clinical adoption hinges on more than just algorithmic accuracy. This qualitative study reveals that healthcare providers are cautiously optimistic about AI-guided dosing but remain deeply concerned about liability, "black box" decision-making, and the potential for technology to erode clinical intuition in complex cases.

What Problem Does This Paper Address?

Traditional medication dosing often follows a "one-dose-fits-all" paradigm, which fails to account for individual variability in how patients process drugs (pharmacokinetics) and how drugs affect their bodies (pharmacodynamics). While "precision medicine" aims to tailor treatments to an individual's genetics and phenotype, the sheer volume of data required to optimize dosing is often too complex for human clinicians to process in real-time.

Artificial Intelligence (AI) offers a solution by identifying ideal dosing specifications through sophisticated data processing. However, even the most accurate tool is useless if clinicians refuse to use it. This paper addresses the "human bottleneck" by exploring the perspectives, fears, and expectations of the healthcare providers (HCPs) who are expected to integrate these tools into their daily workflows.

Methods

The researchers conducted a qualitative study using in-depth, semi-structured interviews with 16 healthcare professionals at Alexandra Hospital in Singapore.

The study utilized an adapted version of the Unified Theory of Acceptance and Use of Technology (UTAUT) framework. While the standard UTAUT model looks at performance expectancy, effort expectancy, social influence, and facilitating conditions, the authors added a critical fifth dimension: Trust and Risk Perception.

During the interviews, participants were introduced to CURATE.AI—an AI dosing tool that uses an individual’s own dose-response data to predict ideal specifications—as a real-world example to ground the discussion. The data was analyzed using a hybrid deductive-inductive coding approach to identify key themes.

Datasets

As this is a qualitative study, the "dataset" consists of the primary source material gathered by the researchers:

  • Interview Transcripts: Verbatim recordings from 16 participants (9 doctors, 4 nurses, and 3 pharmacists).
  • Demographic Data: The cohort represented a mix of specialties (gerontology, cardiology, rheumatology, etc.) and experience levels (69% had >10 years of experience).

The paper does not involve the training of a new machine learning model, but rather evaluates the reception of models like CURATE.AI.

Key Insights and Ideas

  • The "Psychic Partner" vs. The Replacement: HCPs generally don't see AI as a replacement but as a "virtual partner" or "on-call pharmacist." One participant described it as a "psychic partner" that helps make better-informed choices under tight time constraints.
  • The Complexity Paradox: There is a clear divide in where AI is deemed useful. Providers felt AI was perfect for "simple" cases with structured guidelines (like gout or anticoagulation). However, they were skeptical of AI’s ability to handle "complex" cases—particularly elderly patients with multiple comorbidities—where "human instinct" and subjective behavioral assessment are vital.
  • The "Alert Fatigue" Barrier: A major practical concern was the integration into existing Electronic Medical Record (EMR) systems like EPIC. If the AI tool adds to the "barrage of prompts" clinicians already face, it is likely to be ignored or seen as a box-ticking exercise rather than a clinical aid.
  • Erosion of Critical Thinking: A significant "aha moment" in the study was the fear of "over-reliance." Clinicians worried that if they depend too much on AI, they might lose their own critical thinking skills or "lose track of their own learning," potentially leading to disaster if the system fails.

Key Results

  • High Intention to Use: Driven by the belief that AI can enhance productivity, improve task efficiency, and build confidence in medication dosing. Key enablers include a user-friendly interface and institutional endorsement.
  • Low Intention to Use: Driven by concerns over poor reliability, high cognitive load, and "black box" outputs.
  • Liability is the Elephant in the Room: A recurring theme was the lack of clarity regarding legal responsibility. If an AI recommends a dose that harms a patient, who is at fault? Most participants felt that as long as the doctor has to "approve" the medication, the liability remains with the human, which increases their hesitation to trust the tool blindly.
  • Transparency is Non-Negotiable: For AI to be accepted, it must be "explainable." Clinicians need to understand why a specific dose was recommended to foster trust.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Center Study: The research was conducted at one hospital in Singapore, which may limit the generalizability of the findings to other healthcare cultures.
  • Selection Bias: Participants who volunteered might be more "technologically minded" than the average healthcare provider.
  • Researcher Bias: While a standardized topic guide was used, the influence of the interviewers' own perspectives cannot be entirely ruled out.

Future Work: The study calls for real-world implementation evaluations that capture patient outcomes and workflow impacts. There is also a pressing need to develop better Explainable AI (XAI) models that can provide clinicians with the rationale behind recommendations without oversimplifying the clinical reality.

Summary updated

Abstract

Our study highlights the hopeful anticipation and complex challenges of introducing AI-enabled dosing into clinical practice. As AI inevitably becomes a part of healthcare, ongoing evaluation is essential to demonstrate value and promote adoption.
BibTeX
@inproceedings{sumner2025artificial,
  author = {Jennifer Sumner and Jina Ali and Mehul Motani and Abigail Ang and Dean Ho and Amartya Mukhopadhyay},
  title = {{Artificial intelligence guided dosing decisions: a qualitative study on health care provider perspectives}},
  booktitle = {BMJ},
  year = {2025},
  doi = {10.1136/bmjhci-2025-101461},
}
Figure