
Research project
Early Detection of Maternal, Perinatal, and Child Health Risks and their Timely Management using Artificial Intelligence (Maternal-AI).
Maternal health remains a critical public health concern in Tanzania, where maternal mortality rates are among the highest in the world. Contributing factors include limited access to quality healthcare, delayed recognition of high-risk conditions, and resource constraints that challenge timely interventions. Addressing these gaps, the project has implemented an AI-powered tool to support the early detection of maternal complications in collaboration with two public hospitals and three health centers in Dar es Salaam. By utilizing clinical data sourced from existing electronic health records (EHRs) and in-person consultations, the tool is designed to identify critical risks such as anemia, miscarriage, urinary tract infections (UTIs), antepartum complications, pregnancy-induced hypertension (PIH), and postpartum conditions. Initial validation with medical experts has demonstrated the tool’s accuracy, and ongoing efforts are focused on further data collection to improve the model’s performance and generalizability. As the project moves towards real-world pilot testing, it aims for large-scale deployment to enhance maternal, perinatal, and child health outcomes in Tanzania. |
The problem
Problem
The AI tool aims to solve Tanzania's high maternal mortality rate (MMR)
Background
The problem involves four key delays: reaching facilities, seeking care, treatment, and arrival
Consequences
Key complications if not addressed include continued high maternal mortality rates, which conflicts with SDG targets to reduce global MMR to less than 70 per 100k births. Additional challenges include transport issues, absence of health workers, high caseloads, and limited infrastructure/diagnostic ability
Specific issue addressed
It addresses diagnosis and management of maternal health complications including abortion, anemia, antepartum issues, UTI, postpartum complications, and pregnancy-induced hypertension
Signs and indicators
Indicators include various delays in care and treatment
How it is handled today
Traditional handling appears to be manual assessment by healthcare workers
What the tool improves
AI tool aims to improve through early detection and timely management of risks
Scale and context
Global picture
Global target (SDG) is to reduce MMR to less than 70 per 100,000 births
In Tanzania and the region
Specific to Tanzania, which has one of the highest MMR rates
How it works
Purpose
Early detection and management of maternal health risks
Origin
Developed as part of AI4D Africa's Anglophone Multidisciplinary Research Lab
Technology
Uses text-based machine learning/LLM technology
How it works, step by step
Workflow: Login/Register → Input Text → Get Results (2-5s)
Data
Data sources
Patient-oriented attributes (medical history, complaints, etc.)
Diagnostic attributes (doctor's plans, observations, examinations)
Inputs
Patient explanations, doctor's observations, test results
Outputs
Classification of complications (abortion, anaemia, antepartum, UTI, postpartum, pregnancy-induced hypertension)
Who it helps
Primary users
Medical professionals/Doctors
Recommendations it gives
Provides validation based on confidence levels
Efficiency gains
Overall accuracy: 80%
Average sensitivity: 64%
Specific sensitivity rates vary by condition (33%-86%)
Processing time: 2-5 seconds
Evaluation and limitations
Current limitations
Limited demographic scope (Amana Regional Referral Hospital only)
Handling discrepancies
Several technical issues help explain the varying performance levels seen across different maternal health conditions in the AI system. The most significant is data imbalance, where some conditions like pregnancy-induced hypertension have abundant training samples (567) while others like postpartum complications have very few (61). Additionally, the system's reliance on data from a single hospital limits its ability to generalize across different healthcare settings and patient populations.
Performance monitoring
The tool's monitoring framework centers around a 50% confidence threshold for validation, suggesting a balanced approach between sensitivity and specificity
Evaluation questions
Clinical Accuracy and Reliability
- Why do sensitivity rates vary so dramatically across conditions (from 33% for UTI to 86% for postpartum)?
- How does the tool's 80% overall accuracy rate compare to human clinician diagnoses?
Healthcare Delivery Impact
- How does the 2-5 second processing time affect real-world patient care compared to traditional diagnostic approaches?
- Does the tool actually reduce the identified delays (seeking care, reaching facilities, treatment, arrival)?
- Are early referrals based on the tool's recommendations proving appropriate and timely?
Resource Utilization
- Does the tool demonstrably reduce healthcare worker caseload?
- How cost-effective is the implementation compared to traditional care methods?
Equity and Bias
- How does the tool's performance vary across different patient demographics?
- Does the single-hospital training data create biases in predictions?
- Are there specific patient populations for whom the tool is less effective?
System Integration
- How well does the tool integrate into existing hospital workflows?
- Are healthcare workers able to effectively interpret and act on the tool's recommendations?
- Does the confidence level validation system appropriately filter unreliable predictions?
Impact
Practical applications
It enables rapid detection of maternal complications, processes cases in 2-5 seconds, and helps manage high patient caseloads efficiently, By providing quick diagnostic support that enhances healthcare workers' capacity to deliver timely care.
Support for decision-making
The tool offers valuable support for strategic healthcare planning and resource allocation.
Community benefits
Increased satisfaction for clinicians, patients, and community
Proper resource and medication use
Shortened time for management
Improved maternal and child health awareness
Reduced maternal and perinatal death
Scaling opportunities
Yes
Next steps
Long-term goals
Plans to extend project scope
Planned features
Treatment recommendation features planned
Partnerships
Specific partnerships with different health care providers
Access
Demo access
Online Links, not open source
User guidelines
Currently not available




