AI for Healthcare in Low-Resource Settings

Course outline and learning expectations

This is a tutor-led course. The outline shows what your tutor will cover; teaching materials and examinations are provided directly to enrolled students.

What you will learn

• Identify realistic AI opportunities and unsuitable uses within resource-constrained health systems
• Frame a healthcare AI task as screening, prioritisation, education, surveillance or workflow support rather than automatic diagnosis
• Prepare synthetic, de-identified or aggregate health data for responsible model development
• Build and evaluate a symptom-based risk-screening model using sensitivity, specificity, precision, recall, F1-score and calibration concepts
• Develop an introductory medical-image classification workflow and assess image quality, bias and generalisation limits
• Create a simple outbreak-forecasting or anomaly-detection workflow from aggregate reported case data
• Design a health-education or triage chatbot with approved information, red-flag escalation and human referral
• Apply privacy, data minimisation, security, fairness, consent and human-oversight principles to health AI projects
• Work responsibly with small, imbalanced, missing, noisy or changing health datasets
• Build, document and present a malaria risk-screening educational prototype that combines symptom features with aggregate outbreak indicators

Modules

1. 1

   AI Opportunities in Resource-Constrained Health Systems

   Identify where AI may add value and where it may create unacceptable risk

   Health-system workflowsScreening, prioritisation, education and surveillance tasksWorkforce and infrastructure constraintsConnectivity and device limitationsData availability and qualityEquity and accessibilityHuman oversightPrototype versus approved clinical system
2. 2

   Disease-Risk Screening from Symptoms and Basic Clinical Data

   Build and evaluate a classroom risk-screening model without presenting it as a diagnosis

   Problem and label definitionSynthetic or de-identified tabular dataFeature selectionMissing valuesLogistic regression and tree-based baselinesClass imbalanceSensitivity and specificityPrecision, recall and F1-scoreThresholds and calibrationReferral and human-review rules
3. 3

   Image-Based Diagnosis Support Concepts

   Explore medical-image classification using teaching datasets and strict limits on clinical claims

   Malaria microscopy image conceptsSkin-condition teaching imagesImage quality and labellingResizing and normalisationData augmentationCNN and transfer-learning workflowDevice and lighting variationSkin-tone and population biasConfusion matrix and error reviewQualified-reader confirmation
4. 4

   Health Data Privacy, Ethics and Governance

   Apply extra safeguards to sensitive health data and AI-supported decisions

   Uganda Data Protection and Privacy principlesHealth-sector privacy and confidentialityLawful purpose and consentData minimisationDe-identification and re-identification riskAccess control and secure storageFairness and inclusionExplainability and documentationClinical accountabilityApprovals, audit trails and incident response
5. 5

   Predicting Disease Outbreaks from Reported Case Data

   Create a cautious early-warning prototype using aggregate time-series data

   Aggregate case-reporting dataDHIS2-style CSV structureData cleaning and completenessWeekly trendsMoving averagesLag and seasonal featuresBaseline forecastingAnomaly detectionTime-based validationFalse alarms and missed eventsUncertainty communicationPublic-health review and response
6. 6

   Chatbot-Based Health Triage and Education Tools

   Design a low-bandwidth conversational prototype that educates, escalates and protects users

   Defining an approved scopeHealth-education contentIntent and response designRule-based and retrieval approachesRed-flag symptomsEmergency and facility referralLocal-language and literacy considerationsSMS and low-bandwidth interfacesPrivacy-safe loggingTesting unsafe or misleading responsesHuman handover
7. 7

   Responsible Modelling with Limited and Noisy Health Data

   Improve reliability without overstating what a small dataset can prove

   Small samplesMissing and inconsistent recordsLabel errorsRare outcomes and class imbalanceData leakagePatient-level train-test separationCross-validationResampling cautionsRobustness checksDataset documentationModel driftMonitoring and communicating limitations
8. 8

   Capstone — Malaria Risk-Screening Educational Prototype

   Build a non-diagnostic prototype combining symptom features with local aggregate outbreak indicators

   Define users and safety boundariesSelect synthetic, de-identified or aggregate dataPrepare symptom and outbreak featuresTrain a baseline modelChoose a cautious risk thresholdAdd advice to seek approved malaria testingCreate a simple interfaceEvaluate errors and subgroup performanceDocument privacy, bias and clinical limitationsPresent the prototype and human-review workflow