Practical AI in African Healthtech: Where We're Using It Today (And Where We're Not)

AI hype is everywhere in healthtech - diagnosing diseases from images, predicting outcomes, automating clinical workflows. At TIBU Health, we use AI, but not always in the ways the headlines suggest. Here’s what’s actually working for us in an African context, and what remains aspirational.

Where AI is working for us today

1. Appointment no-show prediction

We trained a model on historical data - time of day, demographics, past behaviour, weather - to predict no-show likelihood. High-risk appointments get extra SMS reminders 48 hours and 4 hours before.

Result: No-show rate dropped from 18% to 11%.
Why it works: It solves a narrow, well-defined problem with abundant data and a clear action attached to the prediction.

2. Medication inventory forecasting

We use time-series forecasting to predict stock needs based on prescription patterns and seasonal trends. Alerts trigger automatic reordering.

Result: Stock-outs reduced by 40%.
Why it works: Historical prescription data is relatively clean, and we kept manual override available for pharmacists.

3. Triage prioritisation

A lightweight NLP model flags urgent keywords - chest pain, difficulty breathing - in patient texts or voice notes and bumps them to the front of the virtual queue.

Why it works: It prioritises rather than diagnoses. Clinicians make every actual medical decision.

4. Administrative automation

AI extracts lab results from PDFs, auto-populates forms, and generates visit summaries.

Result: Saves clinicians roughly 30 minutes a day on paperwork.
Why it works: It reduces cognitive load without introducing clinical risk.

Where AI isn’t ready yet (for us)

Diagnostic imaging: Models for TB or diabetic retinopathy work in pilots, but production deployment requires clearer local regulations, much higher clinician trust, and better digital infrastructure - we’re still dealing with film X-rays in many partner clinics.
Predictive diagnosis: Predicting disease before symptoms appear requires massive, clean datasets and genetic/lifestyle data we don’t routinely collect yet.
Clinical chatbots: Liability is too high for AI to give clinical advice. We use chatbots only for booking and FAQs.

What makes AI practical in African healthtech

A few principles that’ve shaped how we approach this:

Start narrow. We solve one workflow pain point well rather than trying to build a general-purpose clinical AI.
Prefer simple models first. Logistic regression often outperforms deep learning when data is limited. We reach for complexity only when simpler approaches plateau.
Build for low connectivity. We run lightweight models on-device or cache predictions locally rather than assuming a live API call will succeed.
Prioritise explainability. Clinicians need to understand why a model flagged a patient as high-risk. A black-box prediction doesn’t build trust.
Human-in-the-loop. AI suggests. Humans decide. This is non-negotiable for us at this stage.