Governance, Accountability and Post-Deployment Monitoring Preferences for AI Integration in West African Clinical Practice: A Mixed-Methods Study

This mixed-methods study of West African clinicians and technical experts reveals a strong preference for independent regulatory oversight, transparent algorithms, and clear accountability frameworks to ensure safe and equitable AI integration in clinical practice, while highlighting significant concerns regarding vendor control and potential unfair liability for medical errors.

The Gist

Imagine you just bought a brand-new, super-smart robot assistant to help you run a busy hospital in West Africa. This robot can diagnose diseases, predict outbreaks, and suggest treatments faster than any human. It sounds like a miracle, right?

But here's the catch: What happens when the robot makes a mistake? Who do you blame? The doctor who used it? The company that built it? Or the government that allowed it? And how do you know the robot is still working correctly tomorrow, next month, or next year?

This paper is essentially a survey of West African doctors and tech experts asking them: "How do we make sure this robot is safe, fair, and accountable before we let it loose in our hospitals?"

Here is the breakdown of their findings, explained with some everyday analogies:

1. The "Who's in Charge?" Problem (Governance)

The Situation: Doctors were asked, "Who should be the referee watching this AI robot?"
The Options:

• The Vendor (The Robot Maker): "We made it, so we'll watch it."
• The Hospital: "We bought it, so we'll watch it."
• The Government: "We are the state, so we'll watch it."
• An Independent Body: A neutral third party, like a sports referee who doesn't work for either team.

The Verdict: The doctors overwhelmingly said, "No way to the Robot Maker!" (Only 3.7% trusted them). They also didn't fully trust the government or the hospital alone.
The Analogy: Imagine a car manufacturer saying, "Don't worry, our cars are safe, just take our word for it." Most people would say, "No thanks, I want an independent safety inspector to check the brakes." The doctors want a neutral referee (40.4% preference) who isn't trying to sell the robot or save the hospital money.

2. The "Live Scoreboard" vs. The "Year-End Report" (Monitoring)

The Situation: How often should we check if the robot is still doing a good job?
The Options:

• Annual Report: "Here is a report on how the robot did last year."
• Real-Time Dashboard: A live screen showing the robot's performance right this second.

The Verdict: Doctors hated waiting for annual reports. They wanted Real-Time Dashboards (41.9% preference).
The Analogy: Think of it like a GPS. You don't want a map that tells you where you were last week. You want a live GPS that screams, "Traffic jam ahead! Reroute now!" If the AI starts giving bad advice because the data has changed (like a new virus strain), doctors want to know immediately, not in a yearly newsletter.

3. The "Blame Game" (Accountability)

The Situation: If the AI gives a wrong diagnosis and a patient gets hurt, who goes to jail or loses their license?
The Fear: The doctors are terrified of being the "fall guy." They fear that if the robot messes up, the hospital will say, "It was the doctor's fault for trusting the robot," even if the doctor followed all the rules.
The Verdict: 76.5% of doctors are worried they will be unfairly blamed.
The Analogy: Imagine you are driving a self-driving car. If the car crashes because of a software glitch, you don't want the police to arrest you for "bad driving." You want the law to clearly say: "If the software failed, the software company is responsible." The doctors are screaming for a clear rulebook that says, "If the AI breaks, the AI maker pays, not the doctor."

4. The "Drift" Problem (Why Continuous Monitoring Matters)

The Situation: AI isn't like a hammer; it doesn't stay the same. It changes over time.
The Analogy: Imagine a weather forecast model trained on data from 2020. If a new, hotter climate pattern emerges in 2026, that old model will start predicting sunny days during a hurricane. This is called "Model Drift."
The Findings: The experts in the study said we need to constantly check if the AI is still "fair" to everyone (rich vs. poor, city vs. village) and if it's still accurate. If it starts drifting, it needs to be retrained, recalibrated, or turned off immediately.

5. The Big Picture: What Do They Want?

The study concludes that West African doctors are ready to use AI, but only if they have a safety net. They don't want to be guinea pigs.

Their "Recipe for Safety" looks like this:

1. An Independent Referee: A neutral group to watch the AI, not the company that built it.
2. Live Cameras: Real-time screens showing if the AI is working correctly right now.
3. A Shield for Doctors: Clear laws that protect doctors from being blamed when the technology fails.
4. A "Stop" Button: A clear plan for what to do if the AI starts acting weird (pause it, fix it, or retire it).

The Bottom Line

This paper is a wake-up call. It says: "We can't just throw fancy new AI tools into our hospitals and hope for the best."

If we don't build these safety nets, doctors will be too scared to use the technology, and patients might get hurt. But if we build a system based on trust, transparency, and clear rules, AI could become the best helper West African healthcare has ever seen. It's about making sure the robot serves the doctor, not the other way around.

Technical Summary

1. Problem Statement

The integration of Artificial Intelligence (AI) into healthcare in West Africa is transitioning from pilot projects to real-world deployment. However, the region lacks established governance frameworks, raising critical concerns regarding:

• Safety and Equity: The risk of "digital divide 2.0," where tools trained on non-local data fail in diverse West African contexts (linguistic, cultural, and epidemiological differences).
• Accountability Gaps: Uncertainty regarding liability when AI systems fail, potentially leading to clinicians being unfairly blamed for algorithmic errors.
• Post-Deployment Risks: Unlike static medical devices, AI models suffer from "data drift" and "concept drift," requiring continuous monitoring to prevent performance degradation.
• Trust Deficits: A lack of empirical data on how West African clinicians and technical experts prefer AI to be governed, monitored, and held accountable.

2. Methodology

The study employed a cross-sectional mixed-methods design conducted between February 22–28, 2026, across West Africa.

• Quantitative Phase:
  • Population: 136 physicians affiliated with the West African College of Physicians (WACP).
  • Sampling: Convenience sampling via WACP communication channels.
  • Instrument: A structured, self-administered questionnaire using a 5-point Likert scale (1 = Strongly Disagree to 5 = Strongly Agree).
  • Data Collection: Conducted via Open Data Kit (ODK) platform; 15–20 minutes duration.
  • Analysis: Descriptive statistics, t-tests, ANOVA, and multivariate linear regression using SPSS v29.
• Qualitative Phase:
  • Population: 72 Key Informant Interviews (KIIs) with technical leads, AI developers, data scientists, policymakers, senior clinicians, and hospital leadership.
  • Instrument: Semi-structured interview guides.
  • Analysis: Thematic analysis using the framework approach (Braun & Clarke, 2021) via NVivo v14.
• Ethical Considerations: Approved by the University of Nigeria Teaching Hospital (UNTH/HREC/2026/01/4157); electronic informed consent obtained; no personally identifiable information collected.

3. Key Contributions

This study provides the first empirical evidence from West Africa regarding clinician and expert preferences for AI governance. Its primary contributions include:

• Defining Governance Preferences: Quantifying the specific trust levels clinicians place in independent bodies vs. vendors or government ministries.
• Identifying Monitoring Needs: Establishing a preference for real-time, transparent data over periodic reporting.
• Highlighting Liability Fears: Documenting the specific fear among clinicians of being scapegoated for AI errors, a critical barrier to adoption.
• Proposing a Framework: Synthesizing quantitative and qualitative data into a actionable, multi-layered governance model tailored for resource-constrained settings.

4. Key Results

A. Quantitative Findings (Survey of 136 Physicians)

• Preferred Oversight Body:
  • Independent Regulatory Bodies: 40.4% (Highest preference; Mean Trust: 4.3/5).
  • Hospital Facilities: 32.4%.
  • Government Ministries: 20.6%.
  • Vendors/Manufacturers: 3.7% (Lowest preference; Mean Trust: 2.4/5).
• Monitoring Preferences:
  • Real-time Dashboards: 41.9% (Most preferred).
  • Monthly Reports: 19.9%.
  • Annual Reports: 1.5% (Least preferred).
• Essential Components:
  • Clear accountability pathways: 94.1% rated as "Essential/Very Important."
  • Algorithm transparency: 91.9%.
  • Real-time performance data: 89.7%.
• Major Concerns:
  • Clinicians being unfairly blamed for AI errors: 76.5%.
  • Excessive vendor control: 72.8%.
  • Lack of clear reporting pathways for harm: 69.9%.

B. Qualitative Findings (72 KIIs)

• Continuous Monitoring: Experts emphasized that deployment is the "starting line," not the finish line. Monitoring must track accuracy, fairness, and safety continuously, specifically stratified by demographics (age, gender, location) to detect localized bias.
• Drift Detection: Systems must automatically detect "data drift" (e.g., new disease outbreaks) and "concept drift" (e.g., changing treatment guidelines) with automated alerts.
• Failure Protocols: A mature framework requires clear steps for when AI underperforms:
  1. Immediate suspension/restriction.
  2. Root-cause analysis.
  3. Recalibration or retirement of the model.
  4. Transparent communication to stakeholders.
• Governance Structure: Recommendations include establishing independent multi-stakeholder bodies, institutional AI committees within hospitals, and strict legal liability frameworks.

5. Significance and Implications

• Policy Roadmap: The study offers a concrete blueprint for West African policymakers to build trustworthy AI frameworks. It argues that without independent oversight and clear liability laws, AI adoption will be hindered by clinician skepticism.
• Addressing the "Responsibility Gap": By highlighting the 76.5% fear of unfair blame, the study underscores the urgent need for legislation that allocates liability to developers or deployers rather than end-user clinicians for algorithmic failures.
• Contextual Relevance: The findings challenge the applicability of high-income country governance models, emphasizing the need for independent (rather than government or vendor-led) oversight in regions with variable regulatory capacity.
• Technical Infrastructure: The strong preference for real-time dashboards suggests that future AI deployment in the region must prioritize infrastructure capable of continuous, stratified performance tracking rather than static, periodic audits.

Conclusion: The study concludes that West African clinicians are ready to engage with AI, but only under governance frameworks that prioritize independent oversight, real-time transparency, and explicit protection against unfair liability. Failure to implement these structures risks undermining the potential benefits of AI and exacerbating health inequities in the region.