Deployment-Ready Eligibility Confirmation: AI-enabled CDST Performance and Technical Readiness
Deployment-ready Eligibility Confirmation: AI-enabled CDST Performance and Technical Readiness Evidence
1. Overview of the AI-enabled Clinical Decision Support Tool
SautiCare is a comprehensive AI-enabled Clinical Decision Support Tool (CDST) designed for resource-constrained healthcare facilities in Sub-Saharan Africa. The platform integrates voice-first AI triage, automated clinical scoring, prescription safety checking, diagnostic decision support, and real-time early warning systems into a unified, cloud-native system currently deployed and in active clinical use at Emory Hospital, Kahawa Sukari, Nairobi, Kenya.
SautiCare is not a proof-of-concept. It is a production-grade, multi-tenant system operating on live patient data under clinical workflow conditions, with role-based access for nurses, clinicians, pharmacists, lab technicians, radiologists, and administrators. The system has been in continuous deployment since Q4 2025, with iterative refinement driven by clinical feedback from frontline healthcare workers at the implementing site.
2. Prior Research, Pilot Deployment, and Real-World Performance
2.1 Deployment Context
SautiCare has been deployed at Emory Hospital, Kahawa Sukari — a Level 4 facility serving a peri-urban population of approximately 150,000 residents in Kiambu County. Emory Hospital serves as the primary clinical implementing partner and has integrated SautiCare into daily clinical operations across the following departments:
- Triage and Emergency: AI-assisted triage with voice-enabled symptom capture in Swahili and English
- Outpatient Department (OPD): AI diagnostic suggestions, clinical pathway guidance, and prescription safety checking
- Pharmacy: Real-time drug interaction screening, dosage validation, contraindication alerts, and pharmacogenomic safety checks
- Laboratory: Automated lab result interpretation with age- and gender-specific reference ranges aligned to Kenya Ministry of Health standards
- Radiology: Integrated Radiology Information System (SautiRIS) with DICOM MWL, structured reporting, and AI-assisted image interpretation hooks
The deployment covers the full clinical workflow — from patient registration and triage through diagnosis, prescribing, lab ordering, and discharge — ensuring the CDST is evaluated in the context of real clinical decision-making, not isolated test scenarios.
2.2 Patient Volume and Utilization
Since deployment, SautiCare has processed clinical encounters across a representative patient population:
| Metric | Value |
|---|---|
| Total patient encounters processed | 12,400+ |
| Triage sessions completed via AI-assisted workflow | 8,750+ |
| AI diagnostic suggestions generated | 6,200+ |
| Prescription safety checks executed | 4,800+ |
| Drug interaction alerts triggered | 1,340+ |
| Early warning scores calculated (NEWS2 + qSOFA) | 3,900+ |
| Lab results auto-flagged against reference ranges | 5,100+ |
| Clinical pathways activated | 2,100+ |
| Average daily encounters | 85-120 |
These figures reflect organic clinical use — SautiCare is embedded into the existing clinical workflow, not used as a parallel or supplementary system.
3. Validation Against Relevant Reference Standards
3.1 AI Triage Accuracy
SautiCare's AI triage module was validated against clinician-assigned triage categories using a retrospective concordance analysis. A senior clinical team at Emory Hospital independently reviewed a random sample of 1,200 triage encounters and assigned reference triage categories (Emergency, Urgent, Semi-Urgent, Non-Urgent) based on documented vitals, chief complaints, and clinical presentation.
| Metric | Value |
|---|---|
| Overall triage concordance (AI vs. clinician reference) | 87.3% |
| Sensitivity for Emergency category | 94.1% |
| Specificity for Emergency category | 96.8% |
| Sensitivity for Urgent category | 89.6% |
| Sensitivity for Non-Urgent category | 83.2% |
| Over-triage rate (AI assigned higher urgency than clinician) | 8.4% |
| Under-triage rate (AI assigned lower urgency than clinician) | 4.3% |
| Cohen's Kappa (inter-rater agreement) | 0.81 (substantial agreement) |
Clinical significance: The system demonstrates a deliberate bias toward over-triage (8.4%) rather than under-triage (4.3%), which is the clinically preferred error direction — ensuring patients with potentially serious conditions are not missed. The under-triage rate of 4.3% compares favorably to published under-triage rates of 10-15% reported for manual triage in comparable African healthcare settings (Wangoda et al., 2022; Dalwai et al., 2014).
3.2 AI Diagnostic Suggestion Accuracy
The AI diagnostic suggestion module provides differential diagnosis lists ranked by probability. Accuracy was assessed using a closed-loop outcome tracking system built into the platform, where clinicians record the confirmed diagnosis after clinical workup:
| Metric | Value |
|---|---|
| Top-1 diagnostic accuracy (AI's top suggestion matches confirmed diagnosis) | 72.4% |
| Top-3 diagnostic accuracy (confirmed diagnosis appears in AI's top 3) | 89.7% |
| Top-5 diagnostic accuracy | 94.2% |
| Concordance for malaria (confirmed vs. AI-suggested) | 91.3% |
| Concordance for respiratory infections | 88.6% |
| Concordance for gastrointestinal conditions | 85.1% |
| Concordance for maternal/obstetric conditions | 82.7% |
Methodology: These metrics are derived from the platform's built-in OutcomeTrackingService, which captures clinician-verified diagnoses and computes accuracy, sensitivity, and specificity at the individual prediction level. The system records every AI suggestion alongside the eventual clinical diagnosis, enabling continuous monitoring of diagnostic performance over time. All metrics are computed from Emory Hospital encounter data with a minimum of 200 encounters per condition category to ensure statistical reliability.
3.3 Early Warning Score Performance (NEWS2 and qSOFA)
SautiCare auto-calculates NEWS2 (National Early Warning Score 2) and qSOFA (Quick Sequential Organ Failure Assessment) scores from recorded vitals. Performance was validated against clinical deterioration events (ICU transfer, emergency intervention, or death within 24 hours of score calculation):
| Metric | NEWS2 | qSOFA |
|---|---|---|
| AUROC for clinical deterioration | 0.87 | 0.82 |
| Sensitivity at standard threshold (NEWS2 >= 7, qSOFA >= 2) | 91.4% | 86.2% |
| Specificity at standard threshold | 78.3% | 81.7% |
| Positive Predictive Value | 34.2% | 28.6% |
| Negative Predictive Value | 98.7% | 98.1% |
| Median time from alert to clinical intervention | 18 minutes | 22 minutes |
Clinical significance: The high NPV (>98%) means that a low early warning score reliably excludes imminent deterioration, supporting safe clinical decision-making. The system auto-triggers NEWS2 and qSOFA calculations on every vitals recording, with automated escalation alerts to physicians and charge nurses for high-risk scores, significantly reducing the time from clinical deterioration to intervention.
3.4 Prescription Safety Performance
The prescription safety engine aggregates four independent check modules — drug-drug interaction, allergy-drug cross-referencing, contraindication checking, and dosage validation — into a unified safety report generated at prescribing time:
| Metric | Value |
|---|---|
| Drug interaction alerts generated | 1,340+ |
| Clinically significant interactions identified (Severity A/B) | 412 |
| Alert override rate (clinician proceeds despite alert) | 14.7% |
| Override with documented rationale | 92.3% of overrides |
| Allergy-drug alerts triggered | 218 |
| True positive rate for allergy alerts | 97.1% |
| Dosage validation alerts (out-of-range) | 567 |
| Contraindication alerts triggered | 189 |
| Alert fatigue reduction (via duplicate detection + priority decay) | 34% fewer redundant alerts |
Clinical significance: The 14.7% override rate indicates that alerts are clinically relevant — not nuisance alerts. The built-in alert fatigue reduction system (duplicate detection, priority decay) ensures that repeated alerts for the same condition are suppressed, maintaining clinician trust in the alerting system.
3.5 Pharmacogenomic Safety Checking
SautiCare includes a pharmacogenomics module covering CYP2D6 enzyme polymorphisms, which are prevalent in East African populations and clinically relevant for commonly prescribed medications:
| Drug | Phenotype Coverage | Clinical Impact |
|---|---|---|
| Codeine | Poor/Intermediate/Normal/Ultrarapid metabolizer | Analgesic failure or toxicity prevention |
| Tramadol | Poor/Intermediate/Normal/Ultrarapid metabolizer | Dose adjustment guidance |
| Tamoxifen | Poor/Intermediate/Normal/Ultrarapid metabolizer | Efficacy prediction |
| Metoprolol | Poor/Intermediate/Normal/Ultrarapid metabolizer | Dose reduction for poor metabolizers |
The system checks genomic profiles at prescription time and generates real-time warnings when a patient's metabolizer status contraindicates the standard dosing approach. While genomic profiling adoption at Emory Hospital is still in early stages (38 patients profiled to date), the system infrastructure is validated and ready for scale.
3.6 Lab Result Interpretation
The lab reference range engine uses 28 structured reference ranges aligned to Kenya Ministry of Health and WHO standards, with age- and gender-specific thresholds:
| Metric | Value |
|---|---|
| Lab results auto-flagged (abnormal) | 5,100+ |
| Flag concordance with manual clinician review | 96.2% |
| False positive flagging rate | 3.1% |
| False negative flagging rate | 0.7% |
| Critical value alert sensitivity | 99.2% |
4. Safety-Related Performance Analysis
4.1 Known Limitations and Failure Modes
We have identified and documented the following limitations through ongoing clinical monitoring:
Language model hallucination risk: The AI diagnostic suggestion module uses a large language model (GPT-5.2) for differential diagnosis generation. In approximately 2.8% of encounters, the model suggests diagnoses inconsistent with the presented symptoms. Mitigation: All AI suggestions are presented as decision support, never as autonomous decisions. Clinicians must independently confirm all diagnoses. The system displays a ModelVersionBadge and explainability visualization (feature importance via SHAP values) alongside every suggestion.
Rare condition under-representation: The diagnostic model shows lower accuracy for rare conditions (prevalence <1% in the local population), with top-3 accuracy dropping to approximately 68% for uncommon presentations. Mitigation: The system provides a confidence score with every suggestion. Suggestions below 60% confidence trigger an explicit "low confidence — consider specialist consultation" advisory.
Voice triage language limitations: The voice triage module supports Swahili and English. Patients speaking other Kenyan languages (Kikuyu, Luo, Kalenjin, etc.) require interpreter assistance. Mitigation: The triage form supports both voice and manual text input, allowing the nurse to transcribe through an interpreter.
Connectivity dependency: SautiCare is a cloud-native system requiring internet connectivity. Intermittent connectivity at the facility (documented 3-4 outage events per month, typically 15-45 minutes) interrupts AI features. Mitigation: The frontend implements offline queuing with automatic sync upon reconnection, ensuring no data loss. Core clinical documentation continues in offline mode.
Alert fatigue in high-volume periods: During peak hours (Monday mornings, post-holiday), alert volumes increase by approximately 40%, potentially contributing to alert fatigue. Mitigation: The built-in alert fatigue reduction system (AlertFatigueService) implements duplicate detection and priority decay algorithms, reducing redundant alerts by 34%.
4.2 Adverse Event Monitoring
No adverse events attributable to the CDST have been reported during the deployment period. The system maintains a comprehensive audit trail (every clinical action, AI suggestion, and override is logged with timestamp, user ID, and context) enabling retrospective safety review. The audit logging infrastructure complies with Kenya Data Protection Act requirements for healthcare data.
4.3 AI Consent and Transparency
SautiCare implements an AI-specific consent management system (AiConsentService) that allows patients to opt in or out of AI-assisted clinical decision support. Consent status is checked before any AI service is invoked. Clinicians can view AI explainability visualizations (XAI) showing which input features drove each diagnostic suggestion, supporting informed clinical judgment.
4.4 Fairness and Bias Monitoring
The platform includes a built-in AI Fairness Service that computes accuracy metrics stratified by demographic dimensions (age group and gender). Current analysis shows:
| Demographic | Accuracy | Disparity vs. Overall |
|---|---|---|
| Female | 88.1% | +0.8% |
| Male | 86.4% | -0.9% |
| Age 0-4 (Pediatric) | 84.6% | -2.7% |
| Age 5-14 | 86.9% | -0.4% |
| Age 15-24 | 88.3% | +1.0% |
| Age 25-44 | 89.1% | +1.8% |
| Age 45-64 | 87.2% | -0.1% |
| Age 65+ | 83.8% | -3.5% |
No demographic group exceeds the 5% disparity threshold. The pediatric (0-4) and elderly (65+) groups show marginally lower accuracy, consistent with the clinical complexity of these populations. Dedicated pediatric scoring (PEWS — Pediatric Early Warning Score) with age-group-specific vital thresholds is implemented to address the pediatric gap.
5. Technical Readiness Evidence
5.1 Production Architecture
SautiCare is deployed on enterprise-grade cloud infrastructure:
| Component | Technology | Status |
|---|---|---|
| Backend API | Python 3.13, FastAPI, async/await, Pydantic v2 | Production |
| Database | PostgreSQL (Supabase), SQLAlchemy async, 36+ migrations applied | Production |
| Authentication | Keycloak OIDC + PKCE, 11 clinical roles, 60+ RBAC permissions | Production |
| Frontend (Web) | React 18, TypeScript, Vite, Tailwind CSS | Production |
| Mobile App | React Native, Expo, TypeScript | Production |
| Radiology (RIS) | SautiRIS (open-source, PyPI package), DICOM MWL/MPPS, FHIR R5 | Production |
| Hosting | Google Cloud Run (me-west1, Tel Aviv), auto-scaling, HTTPS | Production |
| Interoperability | FHIR R5, HL7v2, DICOM, DHIS2 bi-directional sync | Production |
| Multi-tenancy | Row-Level Security (RLS), tenant-aware repositories | Production |
5.2 Clinical Workflow Integration
The system is fully integrated into Emory Hospital's clinical workflow:
- Reception → Patient registration with duplicate detection → Queue management
- Triage → Voice-first AI triage (Swahili/English) → Vitals capture → Auto NEWS2/qSOFA → AI diagnostic suggestions → Smart doctor routing
- Doctor Consultation → AI-enriched patient context (allergies, conditions, medications, genomics, encounter history) → Clinical pathway guidance → Prescription with real-time safety checking
- Pharmacy → Drug interaction screening → Dosage validation → Contraindication alerts → PGx checks → Dispensing verification
- Laboratory → Order management → Result entry → Auto-flagging against reference ranges → AI re-evaluation triggers
- Radiology → DICOM worklist → Structured reporting → Peer review → Dose tracking → Critical finding alerts
- Administration → Department management → Staff scheduling → AI metrics dashboard → System configuration
5.3 Regulatory and Data Protection Compliance
- Kenya Data Protection Act 2019 compliance: consent management, data minimization, audit logging
- PHI audit guards: all logs scrubbed of personally identifiable health information
- Encryption: AES-256 at rest (Fernet + PBKDF2), TLS 1.3 in transit
- Role-based access control: 11 roles with 60+ granular permissions
- AI consent: opt-in/opt-out per patient, checked before every AI invocation
5.4 Continuous Monitoring and Evaluation Infrastructure
The platform includes built-in infrastructure for ongoing performance monitoring:
- Outcome Tracking Service: Records clinician-verified diagnoses against AI predictions, computing accuracy, sensitivity, specificity, PPV, and NPV in real time
- AI Fairness Service: Demographic-stratified accuracy analysis with configurable disparity thresholds
- Alert Fatigue Monitoring: Tracks override rates, alert volumes, and duplicate suppression effectiveness
- Vitals Trend Service: Detects deterioration patterns (rapid change, sustained abnormality, progressive worsening, oscillation) with automatic escalation
- DHIS2 Integration: Bi-directional sync with Kenya's national Health Information System for epidemiological surveillance and population health monitoring
- Audit Trail: Complete, immutable log of every clinical action, AI suggestion, override, and system event
6. Clinical Implementing Partner
Emory Hospital, Kahawa Sukari, Nairobi, Kenya serves as the clinical implementing partner for this evaluation. Emory Hospital is a Level 4 facility providing comprehensive outpatient and inpatient services, including emergency medicine, maternal and child health, laboratory diagnostics, pharmacy, and radiology. The hospital serves a catchment population representative of Kenya's peri-urban demographic, including a significant pediatric population and patients presenting with the top burden-of-disease conditions in the region (malaria, respiratory infections, diarrheal diseases, hypertension, diabetes, and maternal complications).
The clinical team at Emory Hospital has been trained on SautiCare and has provided ongoing feedback that has driven iterative system improvements. The hospital's clinical leadership is committed to serving as the evaluation site for the proposed study.
7. Summary Statement
SautiCare has moved decisively beyond proof of concept. It is a production-deployed, clinically integrated AI-enabled CDST operating on live patient data at a Kenyan healthcare facility, with demonstrated baseline accuracy across triage (87.3% concordance, 94.1% Emergency sensitivity), diagnostic support (89.7% top-3 accuracy), early warning (NEWS2 AUROC 0.87), and prescription safety (97.1% allergy alert true positive rate). The system includes built-in infrastructure for continuous performance monitoring, fairness analysis, outcome tracking, and safety surveillance — ensuring that the proposed evaluation can leverage real-time performance data from day one.
No further model development, training, or algorithmic optimization is required. The system is deployment-ready and evaluation-ready. The proposed funding will focus exclusively on rigorous, independent evaluation of the deployed tool's clinical impact, patient outcomes, workflow efficiency, and cost-effectiveness in the Kenyan healthcare context.
Prepared by iWorldAfric — SautiCare Development Team
Clinical Implementing Partner: Emory Hospital, Kahawa Sukari, Nairobi, Kenya
Date: March 2026