Cognitive AI-Assisted Primary Care Health Delivery: A Pilot Study in Bangladesh

A 2025 pilot study in Bangladesh demonstrates that a cognitive AI system designed to replicate the full clinical workflow, rather than merely providing risk scores, achieved high diagnostic accuracy (94.7%) across 277 patient encounters, suggesting its potential to effectively address physician shortages and enhance primary care delivery.

The Gist

Imagine a world where a doctor is like a master chef in a busy kitchen. In many parts of the world, especially in places like rural Bangladesh, there are far too few chefs for the number of hungry people waiting for a meal. The chefs are exhausted, the lines are long, and many people go hungry or get sick because they can't get help.

This paper is about a pilot study that tested a new kind of "kitchen assistant" called ClinicalAssist. But this isn't just a robot that chops vegetables (which is what most current medical AI does); it's a smart assistant that can actually cook the whole meal from start to finish, under the watchful eye of a human chef.

Here is the story of what happened, broken down into simple terms:

The Problem: The "Chef" Shortage

In countries like Bangladesh, there are very few doctors, especially in the countryside. It's like having one chef for a whole city. Because of this, people often see unqualified "cooks" (unlicensed village doctors) who might give them the wrong medicine. Also, waiting to see a real doctor takes forever.

Most AI tools today are like fortune tellers. They look at a patient and say, "There is a 30% chance this person has a heart attack." That's helpful for planning, but it doesn't actually do anything to help the doctor. The doctor still has to ask all the questions, figure out the diagnosis, write the plan, and fill out the paperwork. The AI hasn't saved the doctor any time.

The Solution: The "Smart Sous-Chef"

The ClinicalAssist system is different. Instead of just guessing a risk score, it acts like a super-smart sous-chef who knows the recipe for almost every illness.

Here is how it works, step-by-step:

1. The Interview (History Taking): The AI asks the patient questions, one by one, just like a detective. It doesn't just ask random things; it asks the next best question based on what the patient just said. It narrows down the possibilities quickly.
2. The Diagnosis: It builds a list of what the patient might have and checks the facts to find the right answer.
3. The Plan: It suggests a treatment plan based on the latest medical rules.
4. The Paperwork: It writes the entire medical report automatically.

The human doctor (the "Head Chef") is still there. They review the AI's work, give the final "okay," and sign off. But the AI does the heavy lifting of gathering information and writing notes.

The Test: A Year in Bangladesh

The researchers tested this system in two places in Bangladesh over one year (2025):

• Site 1: A rural village (Barura).
• Site 2: An industrial area (Comilla) where factory workers live.

They treated 239 unique patients who came in with various health issues. Some came once; some came back for follow-ups. In total, there were 277 visits.

The Results: How Good Was the Assistant?

The results were surprisingly good. Think of it like a student taking a test:

• Overall Score: The AI got the diagnosis right 94.7% of the time.
• Chronic Diseases (Long-term issues): For things like high blood pressure and diabetes, the AI was nearly perfect, getting 98% right. This makes sense because these conditions are stable and the AI just needs to check if the patient is doing okay.
• Acute Care (Sudden sickness): For sudden problems like fevers, infections, or injuries, the AI got 88.9% right. This is harder because symptoms can be confusing (a fever could be flu, dengue, or pneumonia), but getting it right almost 9 out of 10 times is a huge success for a first try.

The "Force Multiplier" Effect:
The biggest win wasn't just the accuracy; it was the time saved. Because the AI asked the questions and wrote the notes, the doctor didn't have to do those tedious tasks. This means one doctor can see many more patients in a day. It's like giving the chef a robot that does all the chopping and plating, so the chef can focus on cooking and serving more people.

The Catch (Limitations)

The study wasn't perfect, and the authors are honest about it:

• Small Sample: They only saw 239 people. To be sure it works for everyone, they need to test it on thousands.
• One Doctor: Only one human doctor supervised the AI. They need to test it with many different doctors to make sure it works for everyone.
• New Places: The system worked well in these specific towns, but it needs to be tested in different types of hospitals and cities.

The Bottom Line

This paper shows that AI doesn't have to be a scary "black box" that just gives a probability score. If we build AI to do the work a doctor does (asking questions, diagnosing, planning, and writing), it can become a powerful tool to solve the shortage of doctors.

In simple terms: ClinicalAssist proved that a smart AI assistant can help a single doctor act like a whole team, saving time and getting more people the care they need. It's a small step, but a very promising one for the future of healthcare in places where doctors are scarce.

Technical Summary

1. Problem Statement

The paper addresses the critical global shortage of primary care physicians, which is particularly acute in low- and middle-income countries (LMICs) like Bangladesh.

• Structural Deficit: Bangladesh faces a severe rural-urban disparity, with rural areas having roughly 21 doctors per 100,000 people compared to 249 in urban centers. Consequently, over 75% of rural patients rely on unqualified village doctors or unregulated drug dispensers.
• Limitations of Current AI: Existing healthcare AI largely focuses on predictive risk-scoring (e.g., predicting sepsis mortality or readmission). The authors argue these tools fail to address the core bottleneck: they do not reduce the time physicians spend on patient encounters. They generate probability estimates but do not assist in history taking, diagnosis formulation, treatment planning, or documentation.
• Metacognitive Gap: Recent studies (cited as a 2026 working paper and a Jan 2025 Nature Communications study) suggest that current Large Language Models (LLMs) lack the "metacognitive capacity" to ask appropriate follow-up questions, update differential diagnoses dynamically, or recognize the limits of their own knowledge, making them unreliable for autonomous clinical reasoning.

2. Methodology

The study reports on a 2025 pilot deployment of ClinicalAssist, an AI system designed to replicate the entire clinical workflow rather than acting as a predictive overlay.

• System Architecture: ClinicalAssist operates through four discrete phases:
  1. Strategic History Taking: Uses a logic engine (not trained on disease prevalence data, but on expert physician weights) to determine the "next best question." It links symptoms to 15,000+ diagnoses and 2,200+ findings, narrowing differentials within ~2 minutes.
  2. Diagnosis Finalization: Revisits findings to confirm or rule out candidates, aiming to reduce anchoring bias.
  3. Treatment Planning: Generates evidence-based treatment plans for physician review (does not prescribe autonomously).
  4. Automated Documentation: Generates complete encounter notes from structured data, eliminating manual dictation.
• Study Design:
  • Type: Prospective observational deployment study.
  • Location: Two sites in Comilla District, Bangladesh (a rural village clinic in Barura and an industrial area clinic in BSCIC).
  • Duration: January – December 2025.
  • Supervision: A single supervising physician (Dr. Nashif Rayhan) validated all AI outputs. The AI collected history and generated notes; the physician verified the final diagnosis and treatment.
  • Data Collection: 239 unique patients, 277 clinical encounters, resulting in 287 diagnostic opportunities. Data was logged in Excel, tracking symptoms, AI output, and a binary accuracy assessment by the physician.

3. Key Contributions

• Workflow Integration vs. Predictive Overlay: The paper introduces a paradigm shift from AI that predicts outcomes to AI that executes the process of care (history, diagnosis, treatment, documentation).
• Metacognitive Simulation: The system is designed to mimic the cognitive process of a seasoned clinician, including asking clarifying questions and auto-correcting based on new data, addressing the "metacognitive" limitations of standard LLMs.
• Field Validation in Resource-Limited Settings: This is the first systematic field deployment of such a system in a live, resource-constrained environment, testing viability outside of controlled US/European settings.
• Independence from Physician Presence: Unlike "ambient scribe" tools that require a doctor to be present to speak, ClinicalAssist can operate independently during data collection, theoretically allowing non-physician providers to deliver physician-level care.

4. Results

The pilot analyzed 287 diagnostic opportunities, excluding 15 records due to software bugs or unlabelled data, leaving 272 valid encounters for analysis.

• Overall Diagnostic Accuracy: 94.7% (272/287 correct).
• Chronic Care Accuracy: 98.0% (200/204 correct).
  • Top conditions: Hypertension (29.0%), Diabetes (24.5%), Asthma, Diabetic Foot Ulcer, Osteoarthritis.
  • The high accuracy here reflects the system's ability to manage structured history and stability assessments for known conditions.
• Acute Care Accuracy: 88.9% (72/81 correct).
  • Top conditions: Scabies (30.6% of acute cases, reflecting the industrial setting), Pneumonia, UTI, Influenza.
  • The lower accuracy is attributed to the complexity of undifferentiated symptom clusters in acute presentations.
• Treatment Plan Accuracy: Among cases with saved plans, 97.7% (216/221) were validated by the physician. Failures were concentrated in rare or non-standard conditions (e.g., burn care, specific dermatological issues).
• False Discovery Rate: For the top 5 high-volume diagnoses (Hypertension, Diabetes, Asthma, Diabetic Foot Ulcer, Osteoarthritis), the false discovery rate was 0.0%.

5. Significance and Implications

• Clinical Force Multiplier: The study demonstrates that AI can function as a "force multiplier" by reducing the time physicians spend per encounter. By automating history taking and documentation (which typically consumes ~50% of physician time), the system increases patient throughput.
• Scalability in LMICs: The high accuracy in chronic disease management (98%) suggests the system is immediately viable for managing the burden of Non-Communicable Diseases (NCDs) in regions with physician shortages.
• Operational Viability: The pilot proved that a workflow-integrated AI system can achieve "clinical-grade" performance in a real-world setting, provided it is supervised and integrated into existing infrastructure.
• Future Directions: The authors call for randomized controlled trials (RCTs) to measure time savings and patient outcomes, multi-site validation across different epidemiological contexts, and formal economic analysis of the efficiency gains.

Conclusion: The paper argues that the future of AI in primary care lies not in risk scores, but in cognitive AI that replicates the full clinical workflow. The Bangladesh pilot provides the first empirical evidence that such a system can achieve high diagnostic accuracy, offering a potential solution to the global primary care workforce crisis.