Detailed Description: Machine Learning-Based Risk Notification for Iron Deficiency
1. Study Overview and Design This is a single-center, prospective, pragmatic randomized controlled trial (pRCT) conducted at China Medical University Hospital (CMUH) in Taiwan. The study aims to evaluate the clinical impact of a machine learning-based clinical decision support tool on iron deficiency detection and related diagnostic behavior in a real-world outpatient setting.
Unlike traditional explanatory trials with restrictive protocols, this pragmatic design integrates the intervention directly into the existing Electronic Health Record (EHR) system.
2. Participant Population
The study includes adult patients (aged 18 years or older) receiving outpatient care in the following departments at CMUH:
1. Division of General Internal Medicine
2. Department of Family Medicine
3. Division of Hematology and Oncology
All eligible outpatient encounters in which a routine complete blood count (CBC) test is performed as part of standard clinical care are included. No additional tests or procedures are required for study participation.
3. AI Intervention and Clinical Workflow
The AI model used in this study estimates the risk of iron deficiency based on routinely available CBC parameters. The randomization and intervention occur automatically within the EHR system after CBC results become available.
Randomization:
Following completion of the CBC, each eligible encounter is automatically assigned, according to a predefined randomization mechanism, to either the Prompt Group or the Control Group. Randomization occurs prior to physician review of laboratory results.
Prompt Group (Intervention):
Physicians are shown an informational "AI Risk Hint," categorizing the patient as either high risk or low risk for iron deficiency, displayed on the laboratory results interface. For patients classified as high risk, the system automatically checks whether iron-related testing has been performed within the previous 30 days. If no recent testing is identified, an informational prompt suggests consideration of iron-related laboratory evaluation.
Control Group:
Physicians review laboratory results according to usual clinical practice, without display of AI-generated risk information or prompts.
Clinical Independence:
The AI system functions solely as a decision support tool. It does not generate orders, mandate testing, or recommend specific treatments. All clinical decisions, including whether to order follow-up iron studies (e.g., ferritin or other iron indices), remain entirely at the discretion of the treating physician.
4. Outcome Measures
4.1 Primary Outcome:
Overall iron deficiency detection rate, defined as the proportion of patients identified as having iron deficiency among all eligible encounters.
4.2 Secondary Outcomes: 4.2.1 Iron deficiency detection rate among patients with anemia The proportion of patients with anemia who are subsequently confirmed to have iron deficiency among all anemic patients included in the study.
4.2.2 Iron deficiency detection rate among patients without anemia The proportion of patients without anemia who are subsequently confirmed to have iron deficiency among all non-anemic patients included in the study.
4.2.3 Rate of ferritin testing for suspected iron deficiency, defined as the proportion of encounters in which ferritin testing is ordered with the indication "to confirm iron deficiency" within 1 month after the CBC report becomes available.
4.2.4 Marginal diagnostic yield of ferritin testing attributable to AI-assisted prompting This measure reflects the average additional number of iron deficiency diagnoses obtained per additional ferritin test attributable to AI-assisted decision support, defined as the difference in confirmed diagnoses between the AI-assisted group and the control group divided by the difference in ferritin tests ordered between the two groups.
4.2.5 Incremental cost-effectiveness of AI-assisted iron deficiency detection This outcome represents the additional cost required to identify one additional case of iron deficiency attributable to AI-assisted decision support.
5. Risk Mitigation and Ethical Considerations This outcome represents the additional cost required to identify one additional case of iron deficiency attributable to AI-assisted decision support, calculated as the difference in ferritin testing costs between the AI-assisted group and the control group divided by the difference in confirmed iron deficiency diagnoses between the two groups.
Data Privacy:
All data processing and analysis are conducted within CMUH's secure internal systems in accordance with institutional policies and local regulatory requirements. No identifiable patient data are transmitted outside the hospital environment.
