Certificate: AI in Clinical Practice

Artificial intelligence is reaching the African bedside with genuine, evidence-graded capability and real, documented risk. This Level II certificate prepares clinicians, nurses, and frontline health professionals, along with students across the health disciplines, to become critical adopters who use AI where the evidence supports it, interpret and override it safely, validate it locally before trusting it, and supervise it under clinical governance.

Across eight weeks and 15 contact hours, you will move from judging what AI can credibly do, such as WHO-recommended TB computer-aided detection and retinal and cervical screening support, through prompt literacy and reading CDSS outputs, local validation and bias checks for your patient population, and workflow redesign that sits AI inside EMRs and clinical accountability. The course closes with an applied workplace project: a defensible plan to adopt, validate, and govern one AI use in your facility, tagged for the Clinical Informatics and AI CEU category.

Course glossary

• Algorithmic bias — systematic, unfair performance differences across patient groups, usually inherited from non-representative or proxy-flawed training data.
• Alert fatigue — desensitisation from excessive or low-value CDSS alerts, leading clinicians to dismiss even important ones.
• Augmentation architecture — deploying AI as triage or second reader inside clinical governance, with human escalation, rather than as a clinician replacement.
• Automation bias — the tendency to accept a computer's suggestion uncritically, risking missed findings or acting on false alarms.
• Capability ladder — the sequenced build of data quality, then statistical capacity, then machine learning.
• Clinical Decision Support System (CDSS) — health IT combining a clinical knowledge base with patient-specific data to generate evidence-based recommendations or alerts.
• Computer-aided detection (CAD) — software reading a medical image to flag or triage abnormality; WHO-recommended for TB screening in people aged 15+.
• Critical-adopter mindset — using AI where evidence supports it while appraising, validating, and supervising it; the disciplined middle between blind trust and blanket rejection.
• Deployment checklist — go/no-go gate: intended use, local validation, governance and escalation, explainability, audit logging, equity monitoring, accountable institution.
• Electronic medical record (EMR) — a real-time, patient-centred digital record at the point of care; the system AI and CDSS attach to.
• Escalation path — the defined next step and referral route triggered by an AI flag, with the hand-off recorded.
• Evidence posture — a label for how ready an AI use is: WHO-recommended, strong-elsewhere-validate-locally, or early-pilot-only.
• Fabrication (hallucination) — fluent but false output from a language model, such as invented citations, doses, or facts.
• Gold standard (reference standard) — the trusted, independent diagnosis a model's outputs are compared against during validation.
• Human-in-the-loop — a governance design in which a named clinician reviews and remains accountable for the AI-assisted decision.
• Intended use — the specific task and population a model was built and evaluated for; using it outside this voids its evidence.
• Local validation — measuring a model's performance on a representative sample of your own population against a gold standard before deployment.
• Out-of-distribution / edge case — a patient or input outside the population and conditions a model was trained and validated on.
• Override — a clinician's documented decision to act against an AI or CDSS recommendation based on clinical reasoning.
• Positive predictive value (PPV) — the chance a flagged patient is truly affected; falls as disease prevalence falls.
• Post-deployment monitoring — ongoing tracking of real-world performance, subgroup outcomes, and drift after a tool is in clinical use.
• Prompt literacy — the skill of giving clinical AI clear, specific, bounded instructions and critically reading the output.
• Regulatory sandbox — a supervised environment where a new AI tool is trialled and evidence and regulation co-develop before full approval.
• Sensitivity — the proportion of truly affected patients a tool correctly flags; high sensitivity minimises missed cases.
• SMART guidelines — WHO's machine-readable clinical recommendations that deploy national protocols into digital systems at software speed.
• Software as a medical device (SaMD) — software performing a medical function and regulated as a device; pathways now extend to algorithmic tools.
• Specificity — the proportion of truly unaffected patients a tool correctly clears; low specificity produces many false alarms.
• Trust calibration — matching reliance on a tool to its validated performance on your population for its intended use.
• WHO six principles — autonomy; well-being and safety; transparency and explainability; responsibility and accountability; inclusiveness and equity; responsive sustainable AI.

Frequently asked questions

Q: Will AI replace clinicians in African health systems? A: No. Where the baseline is roughly one specialist for millions of people, AI is not replacing workforce — it is creating capacity that would not otherwise exist. The book's posture is "augmentation, not replacement": AI as triage or second reader inside clinical governance, with humans in the loop and a named clinician accountable for every decision.

Q: Which AI uses are actually proven enough to trust in clinical practice today? A: The strongest evidence is for narrow, image- or signal-based tasks where specialists are scarcest. WHO recommends TB computer-aided detection (CAD) on chest X-rays for screening and triage in people aged 15+. Diabetic-retinopathy, cervical-cancer, and dermatology screening support, plus SMART-guideline CDSS, have strong evidence but should be validated locally. Conversational AI is promising but must be governed per WHO's 2024 guidance.

Q: A model worked well in a published trial elsewhere. Why can't I just use it? A: Performance is population- and task-specific. A model trained on other populations can fail silently and differentially on yours — the Obermeyer algorithm and skin-tone-biased dermatology models are documented examples. You must validate it locally on a representative sample against a gold standard, break performance down by subgroup, and run it through the deployment checklist before clinical use.

Q: When am I allowed to override a CDSS or AI recommendation? A: Whenever the patient or input is outside the model's validated scope (an edge or out-of-distribution case), the input is poor quality, or the recommendation contradicts strong clinical findings or the national guideline or formulary. Override is the system working as designed — but always document what the AI said, that you disagreed, and your clinical reasoning, and escalate high-stakes disagreements for a second opinion.

Q: Is it safe to use a chatbot like a large language model for clinical questions? A: Use it for drafts — referral letters, summaries, plain-language explanations — with clear, bounded prompts, and always verify clinical facts against an authoritative source, because LLMs fabricate fluently. Never paste identifiable patient data into a consumer tool you do not control; that breaches confidentiality and data-protection law (Ethiopia's PDPP, POPIA, Nigeria's and Kenya's DPAs). Prefer governed institutional tools and design for low bandwidth.

Q: What is "automation bias" and why does it matter to me? A: It is the human tendency to accept a computer's suggestion uncritically, especially when busy, tired, or junior. It causes both missed findings (trusting a false "normal") and acting on false alarms. The countermeasure is trust calibration — relying on a tool in proportion to its validated performance for its intended use — and keeping the clinician as the appraising decision-maker.

Q: Who is responsible if an AI tool causes harm? A: A named clinician and institution — never the algorithm. Under the WHO principle of responsibility and accountability, a human owns every AI-assisted decision and documents the reasoning. Governance makes this concrete through defined intended use, escalation paths, audit logging, and clear accountability.

Q: How does this course relate to my CPD and the wider regulatory picture? A: It is a Level II, 15-contact-hour certificate tagged for the Clinical Informatics / AI CEU category, delivered as a cohort with an applied workplace project. It aligns with the workforce competency of "critical supervision of algorithmic colleagues" (Chapter 3) and with the consolidating policy layer — the WHO six principles, the African Union's 2024 continental AI strategy, and national software-as-a-medical-device pathways and regulatory sandboxes.