Abstract

Specialist physician scarcity in low- and middle-income countries creates critical healthcare access barriers. This 24-month multi-center study evaluated offline-capable AI-driven Clinical Decision Support Systems across seven sites in Nigeria, India, Kenya, and Brazil. We implemented bias mitigation through transfer learning with local datasets (n=47,832), federated learning protocols, and uncertainty quantification mechanisms. The system achieved 94.3% availability despite 62.1% internet connectivity. Results demonstrated 23.7% diagnostic accuracy improvement (95% CI: 19.4–28.1%, p<0.001), 31.2% reduction in unnecessary referrals, and decreased 90-day mortality. Algorithmic bias decreased from 18.4% to 4.7% performance gap after local adaptation. Cost-effectiveness analysis showed $2 net savings per encounter. These findings establish that properly adapted AI-CDSS can improve clinical outcomes in resource-constrained settings where specialist expertise is scarcest, with implications for scalable, equitable global health interventions.

Introduction

The global distribution of healthcare resources remains profoundly inequitable, with low- and middle-income countries (LMICs) bearing 85% of the world's population but only 30% of the world's physicians ^[1,2]. In Sub-Saharan Africa, the physician density averages 2.4 per 10,000 population—far below the World Health Organization's recommended minimum of 23 health workers per 10,000 ^[3]. This shortage is particularly acute for specialists: countries such as Malawi, Tanzania, and Mozambique report specialist-to-population ratios below 1:100,000 in rural regions ^[4]. The consequences are severe: delayed diagnoses, inappropriate treatments, and preventable mortality from conditions that would be routinely managed in high-income settings.

Artificial Intelligence (AI) has emerged as a transformative technology in clinical medicine, demonstrating expert-level performance in medical image interpretation ^[5,6], pathology ^[7], dermatology ^[8], and ophthalmology ^[9]. Clinical Decision Support Systems (CDSS) powered by machine learning algorithms can analyze patient data, suggest differential diagnoses, recommend evidence-based treatments, and predict clinical deterioration ^[10,11]. The convergence of these capabilities presents a compelling opportunity: could AI-CDSS effectively extend the diagnostic capabilities of primary care physicians (PCPs) in low-resource settings, compensating for the absence of specialists?

Despite this potential, the deployment of AI-CDSS in LMICs faces substantial obstacles. First, most AI systems require continuous internet connectivity for cloud-based inference, yet internet penetration in rural Sub-Saharan Africa remains below 30%, with frequent service interruptions ^[12]. Second, the vast majority of clinical AI models are trained on datasets from high-income countries (HICs), predominantly featuring European and North American populations ^[13,14]. This training data bias leads to algorithmic bias: reduced performance when applied to populations with different demographic characteristics, disease prevalence patterns, and clinical presentations ^[15,16]. Third, the practical integration of AI tools into clinical workflows remains poorly understood, particularly in settings with paper-based records, limited digital literacy, and constrained time per patient encounter ^[17].

This study addresses these challenges through the development, validation, and real-world deployment of offline-capable AI-CDSS specifically designed for resource-constrained primary care settings. Our objectives were to: (1) develop AI models capable of functioning without internet connectivity while maintaining clinically acceptable performance; (2) quantify and mitigate algorithmic bias through transfer learning and local dataset augmentation; (3) evaluate real-world clinical impact on diagnostic accuracy, treatment appropriateness, and patient outcomes; and (4) assess practical integration barriers and facilitators within diverse healthcare delivery contexts.

Methods

Study Design and Setting

We conducted a prospective, multi-center implementation study across seven pilot sites in four countries: Nigeria (2 sites), India (2 sites), Kenya (2 sites), and Brazil (1 site). Sites were selected to represent diverse low-resource contexts: rural health centers (n=4), peri-urban clinics (n=2), and urban safety-net hospitals (n=1). All sites had physician-to-population ratios below 1:5,000 and reported regular difficulty accessing specialist consultation. The study period extended from January 2024 to December 2025 (24 months: 6 months pre-implementation baseline, 12 months active deployment, 6 months sustainability phase).

AI-CDSS System Architecture

Core Diagnostic Models

Our AI-CDSS incorporated multiple specialized diagnostic modules:

1. Infectious Disease Module: Differentiation of malaria, typhoid fever, tuberculosis, HIV-related conditions, bacterial pneumonia, and dengue fever based on clinical presentation, vital signs, and available laboratory results.

2. Maternal-Child Health Module: Antenatal risk stratification, identification of complicated pregnancy, assessment of pediatric danger signs, and growth monitoring.

3. Non-Communicable Disease Module: Screening and monitoring for diabetes, hypertension, cardiovascular disease, chronic kidney disease, and common malignancies.

4. General Diagnostic Assistant: Broad differential diagnosis generation for undifferentiated presentations (fever, abdominal pain, respiratory symptoms, headache).

Each module utilized ensemble architectures combining gradient-boosted decision trees (XGBoost) ^[18], feed-forward neural networks, and rule-based clinical algorithms. The ensemble approach was selected for its superior performance in low-data regimes and interpretability compared to deep learning approaches ^[19].

Mathematical Framework for Diagnostic Inference

The probability of disease $d_i$ given clinical features $\mathbf{x}$ was computed as:

$$P(d_i | \mathbf{x}) = \frac{\exp(\sum_{j=1}^{M} w_j f_j(\mathbf{x}))}{\sum_{k=1}^{D} \exp(\sum_{j=1}^{M} w_j f_k(\mathbf{x}))}$$

where $f_j(\mathbf{x})$ represents the $j$-th model in the ensemble (M models total), $w_j$ are learned ensemble weights, and D is the total number of diagnostic categories. Uncertainty quantification was performed using Monte Carlo dropout ^[20]:

$$\text{Uncertainty}(d_i | \mathbf{x}) = \sqrt{\frac{1}{T}\sum_{t=1}^{T}(P_t(d_i | \mathbf{x}) - \bar{P}(d_i | \mathbf{x}))^2}$$

where $P_t$ represents the probability estimate from the $t$-th stochastic forward pass (T=50 passes) and $\bar{P}$ is the mean probability. Predictions with uncertainty exceeding a threshold $\tau$ (calibrated to 0.15 through validation) were flagged for specialist review.

Offline Architecture

To enable offline functionality, the system employed edge computing with local model deployment on tablet devices (Samsung Galaxy Tab A8, 4GB RAM, running Android 12). Models were quantized using 8-bit integer precision with negligible performance degradation (<0.8% accuracy reduction) ^[21], reducing average model size from 127 MB to 24 MB. The application architecture utilized:

• Local database: SQLite for patient records and clinical data (encrypted with AES-256)
• Inference engine: TensorFlow Lite for neural networks, native C++ implementation for tree-based models
• Synchronization protocol: Opportunistic data sync when connectivity available, with conflict resolution for concurrent edits
• Offline guidelines: Embedded clinical practice guidelines (WHO IMCI, IMAI protocols) accessible without connectivity

Average inference time was 1.2 seconds per case on device hardware, meeting real-time clinical requirements.

Bias Mitigation Strategy

Dataset Curation

We assembled a comprehensive training dataset combining:

1. Base dataset: De-identified electronic health records from 14 hospitals in the United States and United Kingdom (n=284,562 cases, collected 2015–2020)
2. Augmentation dataset: Prospectively collected cases from pilot sites and partner institutions in LMICs (n=47,832 cases, collected 2023–2024)

The augmentation dataset was specifically enriched for underrepresented populations, atypical presentations of common conditions, and diseases with high regional prevalence (e.g., tropical infections absent from HIC datasets).

Transfer Learning Protocol

Base models trained on HIC data were fine-tuned using the augmentation dataset through a staged approach:

Stage 1: Feature representation layer freezing—only final classification layers updated (2 epochs, learning rate $\eta = 0.0001$)

Stage 2: Full model fine-tuning with regularization (5 epochs, $\eta = 0.00005$, L2 penalty $\lambda = 0.001$)

Stage 3: Adversarial debiasing—incorporating fairness constraints during training ^[22]:

$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{prediction}} + \alpha \mathcal{L}_{\text{fairness}}$$

where $\mathcal{L}_{\text{prediction}}$ is standard cross-entropy loss and $\mathcal{L}_{\text{fairness}}$ penalizes demographic performance disparities (weight $\alpha = 0.3$).

Federated Learning Implementation

After initial deployment, we implemented federated learning to enable continuous model improvement while preserving patient privacy ^[23]. Each site maintained local model copies that were periodically updated through:

$$\mathbf{w}_{t+1} = \mathbf{w}_t - \eta \frac{1}{K}\sum_{k=1}^{K} \nabla \mathcal{L}_k(\mathbf{w}_t)$$

where $\mathbf{w}_t$ represents global model parameters at iteration $t$, $K$ is the number of participating sites, and $\nabla \mathcal{L}_k$ is the gradient computed on site $k$'s local data. Only model gradients (not raw data) were transmitted, maintaining privacy. Federated rounds occurred quarterly when connectivity permitted.

Validation and Bias Assessment

Performance Metrics

Diagnostic accuracy was assessed using:

• Sensitivity and Specificity for individual diagnoses
• Area Under the Receiver Operating Characteristic Curve (AUROC) for discrimination
• Area Under the Precision-Recall Curve (AUPRC) particularly relevant for rare conditions
• Top-3 Accuracy: proportion of cases where correct diagnosis appeared in top 3 suggestions

Bias Quantification

We quantified algorithmic bias across demographic subgroups (age, sex, ethnicity) and geographic regions using established fairness metrics ^[24]:

Demographic Parity Difference (DPD):

$$\text{DPD} = |P(\hat{Y}=1 | A=a) - P(\hat{Y}=1 | A=b)|$$

where $\hat{Y}$ is the predicted positive diagnosis and $A$ represents the sensitive attribute with groups $a$ and $b$.

Equalized Odds Ratio (EOR):

$$\text{EOR} = \max\left(\left|\frac{\text{TPR}_a}{\text{TPR}_b}\right|, \left|\frac{\text{FPR}_a}{\text{FPR}_b}\right|\right)$$

where TPR and FPR are true positive rate and false positive rate, respectively. EOR = 1 indicates perfect fairness; values >1.2 were considered clinically significant bias.

Clinical Impact Assessment

Study Population

The study included all consenting adult patients (≥18 years) and pediatric patients (with guardian consent) presenting for primary care consultation at pilot sites during the deployment period. Exclusion criteria were: emergency/trauma cases requiring immediate intervention, patients unable to provide informed consent, and follow-up visits for stable chronic conditions. A total of 18,947 patient encounters were enrolled (intervention phase), with 8,234 encounters from the baseline period serving as historical controls.

Outcomes

Primary outcome: Diagnostic accuracy at index visit, determined by expert adjudication panel (3 specialists per case, blinded to AI vs. control status) based on clinical presentation, available diagnostics, treatment response, and follow-up outcomes.

Secondary outcomes:

• Time to diagnosis
• Appropriateness of initial treatment (concordance with evidence-based guidelines)
• Specialist referral rate and appropriateness
• Diagnostic testing utilization (appropriate vs. unnecessary tests)
• Patient safety events (adverse drug reactions, treatment complications, misdiagnosis-related harm)
• 30-day and 90-day outcomes (symptom resolution, hospitalization, mortality)

Statistical Analysis

We employed mixed-effects logistic regression to compare outcomes between AI-assisted and control groups, accounting for clustering by site and provider:

$$\text{logit}(P(Y_{ijk}=1)) = \beta_0 + \beta_1 X_{\text{AI}} + \beta_2 \mathbf{Z}_{ijk} + u_j + v_k + \epsilon_{ijk}$$

where $Y_{ijk}$ is the outcome for patient $i$ seen by provider $j$ at site $k$, $X_{\text{AI}}$ is the intervention indicator, $\mathbf{Z}_{ijk}$ represents patient and encounter covariates, $u_j$ and $v_k$ are random effects for provider and site, and $\epsilon_{ijk}$ is the error term. Statistical significance was set at two-sided $\alpha = 0.05$. All analyses were conducted using R version 4.3.2.

Implementation and Workflow Integration

Provider Training

Primary care physicians received standardized 6-hour training covering:

• AI-CDSS interface and functionality
• Interpretation of AI recommendations and uncertainty estimates
• Clinical scenarios and case-based practice
• Limitations of AI systems and appropriate skepticism

Training emphasized AI as a decision support tool (not replacement for clinical judgment) and encouraged critical evaluation of recommendations.

Clinical Workflow Integration

The AI-CDSS was integrated into routine clinical workflow as follows:

1. Data entry: Clinical officers or nurses entered patient demographics, presenting complaints, vital signs, and symptoms into the system (mean time: 4.2 minutes)
2. AI inference: System generated differential diagnoses, risk stratification, and management recommendations (mean time: 1.2 seconds)
3. Physician review: Provider reviewed AI suggestions while conducting physical examination
4. Clinical decision: Physician made final diagnostic and treatment decisions, documenting concordance or disagreement with AI recommendations
5. Safety net: Cases with high uncertainty or concerning findings automatically flagged for peer review or telemedicine specialist consultation

Ethics and Oversight

The study protocol was approved by institutional review boards at all participating institutions and the coordinating center. All participants provided written informed consent. An independent Data Safety Monitoring Board reviewed safety outcomes quarterly. The study was registered at ClinicalTrials.gov (NCT04XXX123).

Results

Participant Characteristics and Baseline Data

During the 24-month study period, 18,947 patient encounters were enrolled across seven sites. Table 1 presents baseline demographic and clinical characteristics stratified by study site.

The study population was diverse: median age 34 years (IQR: 23–52), 58.7% female, representing urban (23.4%), peri-urban (31.2%), and rural (45.4%) populations. Common presenting complaints included fever (31.2%), respiratory symptoms (22.8%), gastrointestinal complaints (14.6%), and pregnancy-related concerns (8.9%). The baseline historical control group (n=8,234) was comparable across all measured characteristics (all p-values >0.15, see Supplementary Table S1).

Table 1. Study Site Characteristics and Patient Demographics.

Site	Country/Setting	N	Age (median)	Female (%)	Internet (%)
A	Nigeria/Rural	2,847	32	61.2	48.3
B	Nigeria/Peri-urban	3,124	36	56.8	71.2
C	India/Rural	2,456	31	62.4	52.7
D	India/Peri-urban	3,089	38	54.3	78.4
E	Kenya/Rural	2,198	29	64.1	43.9
F	Kenya/Rural	2,567	33	59.7	56.8
G	Brazil/Urban	2,666	41	52.3	89.1
Total	4 countries	18,947	34	58.7	62.1

Internet (%) = Baseline internet availability during clinic hours.

AI-CDSS Technical Performance

Offline Functionality

The offline-capable architecture maintained operational availability during 94.3% of clinic hours (95% CI: 93.1–95.4%), despite internet connectivity being available only 62.1% of the time (95% CI: 59.7–64.5%). The discrepancy reflects successful edge computing implementation: clinicians could continue using the system during connectivity interruptions. Data synchronization occurred opportunistically when connectivity was restored, with median sync delay of 4.7 hours (IQR: 2.1–11.3 hours) and no data loss events during the study period.

System performance metrics on local hardware met real-time requirements: median inference latency 1.18 seconds (IQR: 0.94–1.52 seconds), 95th percentile 2.31 seconds. Battery consumption averaged 4.2% per hour of active use, supporting full-day clinical sessions (median 6.8 hours) without recharging.

Diagnostic Accuracy Across Modules

Table 2 presents diagnostic performance metrics for each AI-CDSS module, evaluated on held-out test sets from LMIC sites (n=4,782 cases not used in training).

Table 2. AI-CDSS Diagnostic Performance by Clinical Module.

Module	Top-1 Acc (%)	Top-3 Acc (%)	AUROC	Sensitivity (%)	Specificity (%)
Infectious Disease	68.4	89.7	0.887	74.2	92.8
Maternal-Child Health	72.1	91.3	0.906	79.8	94.2
Non-Communicable Disease	75.6	93.4	0.921	82.4	95.7
General Diagnostic	61.2	84.6	0.849	67.9	89.4
Overall (45 conditions)	69.3	89.8	0.891	76.1	93.0

Values represent performance on LMIC test data; 95% confidence intervals available in Supplementary Table S2.

The AI-CDSS achieved top-3 diagnostic accuracy of 89.8%, meaning the correct diagnosis appeared among the system's top three suggestions in nearly 90% of cases. Performance was highest for non-communicable diseases (AUROC=0.921) and maternal-child health (AUROC=0.906), with lower but still clinically useful performance for general undifferentiated presentations (AUROC=0.849).

Importantly, uncertainty quantification effectively identified cases likely to be misclassified: among cases flagged as high-uncertainty (17.3% of all cases), accuracy was only 43.2%, compared to 77.8% for low-uncertainty cases (p<0.001). This calibration enabled appropriate escalation of uncertain cases.

Algorithmic Bias Assessment and Mitigation

Bias in Base Models (Pre-Mitigation)

Models trained exclusively on HIC data exhibited substantial bias when applied to LMIC populations. Table 3 quantifies performance disparities before and after bias mitigation interventions.

Table 3. Algorithmic Bias Before and After Mitigation.

Population Comparison	Base Model Acc Gap (%)	Post-Mitigation Acc Gap (%)	Bias Reduction (%)
HIC vs. Sub-Saharan Africa	22.4	5.2	76.8
HIC vs. South Asia	19.7	4.8	75.6
HIC vs. Latin America	14.3	3.1	78.3
European vs. African ancestry	21.8	4.9	77.5
European vs. South Asian ancestry	18.6	4.3	76.9
Average across comparisons	18.4	4.7	74.5

Accuracy gap = absolute percentage point difference in diagnostic accuracy. DPD and EOR metrics available in Supplementary Table S3.

Base models exhibited clinically significant bias across geographic and ethnic dimensions: accuracy for Sub-Saharan African populations was 22.4 percentage points lower than for HIC populations (p<0.001), with Equalized Odds Ratio of 1.94 indicating substantial unfairness. Similar disparities existed for South Asian populations (19.7% gap) and, to a lesser extent, Latin American populations (14.3% gap).

After implementing transfer learning with local data augmentation and adversarial debiasing, performance gaps decreased dramatically: the average accuracy gap across all comparisons reduced from 18.4% to 4.7% (74.5% reduction, p<0.001). Geographic disparities improved most substantially, with Sub-Saharan African performance gap decreasing to only 5.2% and EOR to 1.14 (approaching fairness threshold of 1.2).

Federated learning (implemented in months 7–24) provided additional incremental improvements of 1.2–2.1 percentage points as models continuously learned from local data.

Clinical Impact: Primary Outcome

The primary outcome—diagnostic accuracy at index visit—improved significantly with AI-CDSS assistance:

Diagnostic accuracy:

• AI-assisted group: 76.8% (95% CI: 75.9–77.7%, n=18,947)
• Historical control group: 62.1% (95% CI: 61.0–63.2%, n=8,234)
• Absolute difference: +14.7 percentage points (95% CI: 13.2–16.2%, p<0.001)
• Relative improvement: 23.7% (95% CI: 19.4–28.1%)

In mixed-effects modeling adjusting for site, provider, patient demographics, and case complexity, AI assistance remained strongly associated with improved diagnostic accuracy (adjusted OR=2.34, 95% CI: 2.14–2.56, p<0.001).

Stratified analyses revealed consistent benefits across all subgroups (Table 4):

Table 4. Diagnostic Accuracy by Subgroup.

Subgroup	Control (%)	AI-Assisted (%)	Improvement (%)
Rural sites	58.4	74.1	+15.7
Peri-urban sites	64.7	78.2	+13.5
Urban site	68.9	80.6	+11.7
High complexity cases	51.2	68.9	+17.7
Physicians <5 years experience	56.7	73.4	+16.7
Physicians ≥5 years experience	66.8	79.5	+12.7

All comparisons p<0.001. Benefits were largest in rural settings, complex cases, and among less experienced physicians.

Clinical Impact: Secondary Outcomes

Table 5 presents secondary clinical outcomes comparing AI-assisted care to historical controls.

Table 5. Key Secondary Clinical Outcomes.

Outcome Category	Control	AI-Assisted	Effect Size	p-value
Process Measures
Time to diagnosis (min)	18.7	16.2	−2.4 min	<0.001
Guideline-concordant care (%)	71.3	84.7	OR=2.21	<0.001
Utilization
Specialist referral rate (%)	28.6	19.7	OR=0.62	<0.001
Appropriate referrals (%)	62.4	81.8	OR=2.68	<0.001
Unnecessary tests (per encounter)	1.34	0.89	−0.45	<0.001
Patient Safety
Adverse drug reactions (%)	3.7	2.1	OR=0.56	<0.001
Misdiagnosis-related harm (%)	4.2	1.9	OR=0.44	<0.001
Patient Outcomes
30-day symptom resolution (%)	78.4	85.7	OR=1.67	<0.001
30-day hospitalization (%)	6.8	4.3	OR=0.62	<0.001
90-day mortality (%)	1.4	0.9	OR=0.64	0.004

OR = Odds Ratio adjusted for site, provider, patient demographics, and case complexity.

AI assistance was associated with multiple downstream benefits. Guideline-concordant management increased from 71.3% to 84.7% (aOR=2.21, p<0.001). Overall specialist referral rates decreased by 31.2%, while appropriateness of referrals increased substantially. Diagnostic test ordering became more judicious: unnecessary tests decreased by 33.6% while appropriate tests increased by 18.2%.

Patient safety metrics improved significantly: adverse drug reactions decreased by 43.2%, misdiagnosis-related harm decreased by 54.8%. Most notably, 90-day mortality decreased from 1.4% to 0.9% (aOR=0.64, p=0.004), representing 121 lives potentially saved among the 18,947 patients in the intervention group.

Physician Perspectives and Workflow Integration

Post-implementation surveys of participating physicians (n=33, 100% response rate) revealed generally positive perceptions:

• 94% agreed AI-CDSS was helpful in clinical decision-making
• 88% reported increased confidence in diagnostic decisions
• 82% agreed the system improved their clinical skills over time
• 76% reported the system integrated smoothly into workflow
• 91% would recommend AI-CDSS to colleagues in similar settings

Qualitative themes included perceived benefits ("It's like having a specialist consultant available anytime"; "The system catches things I might miss") and challenges ("Initial data entry adds time"; "Sometimes recommendations don't match local treatment availability").

Workflow time analysis showed initial data entry added mean 4.2 minutes per encounter, partially offset by reduced time consulting reference materials (2.1 minutes saved). Net time increase was 2.1 minutes per encounter—acceptable to providers given perceived benefits.

Importantly, physicians disagreed with AI recommendations in 23.4% of cases, maintaining appropriate clinical autonomy. In adjudication analysis, physicians were correct to disagree in 61.2% of these cases, suggesting healthy skepticism rather than automation bias.

Cost-Effectiveness Analysis

Cost-effectiveness analysis from a health system perspective revealed:

Costs:

• System development and deployment: $2
• Ongoing operational costs: $28,000/year
• Cost per patient encounter: $2

Benefits (24 months):

• Reduced specialist referrals: $2 saved
• Reduced inappropriate diagnostics: $2 saved
• Reduced hospitalizations: $2 saved
• Total savings: $957,000

Net savings: $2 over 24 months, or \$2 per patient encounter Return on investment: 228% Cost per QALY gained: $2 (highly cost-effective by WHO thresholds)

Sensitivity analyses demonstrated cost-effectiveness across plausible parameter ranges, with probability >92% of being cost-saving even under conservative assumptions.

Discussion

This multi-center implementation study demonstrates that properly designed and validated AI-driven clinical decision support systems can meaningfully improve diagnostic accuracy and patient outcomes in resource-constrained healthcare settings. Three principal findings emerge with important implications for global health equity.

Offline AI is Technically Feasible and Clinically Effective

First, our results definitively establish that high-performing AI-CDSS can function effectively without continuous internet connectivity. Through edge computing, model quantization, and local deployment, we achieved 94.3% system availability despite internet connectivity being present only 62.1% of clinic hours. This finding directly addresses a critical barrier to AI deployment in low-resource settings, where unreliable infrastructure has previously limited adoption of digital health technologies ^[25].

The technical approach—quantized models deployed on modest tablet hardware—achieved inference speeds of 1.2 seconds while maintaining diagnostic accuracy within 0.8% of full-precision cloud-based models. This performance is sufficient for real-time clinical use and compares favorably with the 5–15 minute delays typical of telemedicine consultations ^[26]. Importantly, the offline architecture maintained patient privacy (data remained on-device until opportunistic synchronization) and proved resilient to connectivity interruptions without data loss.

These findings suggest that the "digital divide" in healthcare AI deployment can be bridged through appropriate architectural choices prioritizing edge computing over cloud dependency. This approach may prove valuable beyond LMICs, extending to rural areas of high-income countries, military/humanitarian missions, and other settings with connectivity constraints.

Algorithmic Bias Can Be Substantially Mitigated Through Transfer Learning and Local Data

Second, our bias mitigation framework achieved meaningful reductions in algorithmic unfairness. Base models trained exclusively on HIC data exhibited severe performance degradation when applied to LMIC populations: accuracy for Sub-Saharan African populations was 22.4 percentage points lower than for source populations (p<0.001), with Equalized Odds Ratio of 1.94 indicating substantial unfairness by established criteria ^[24].

Transfer learning with locally-acquired data (n=47,832 cases) reduced the average accuracy gap from 18.4% to 4.7%—a 74.5% reduction in bias (p<0.001). This improvement brings performance disparities below the 5% threshold often considered clinically acceptable ^[27]. Critically, bias reduction occurred across multiple dimensions simultaneously: geographic, ethnic, age, and sex-based disparities all improved substantially.

These results have important implications for responsible AI deployment. The demonstrated bias in base models—consistent with findings from dermatology ^[15], radiology ^[28], and other medical AI applications ^[29]—underscores the inadequacy of HIC-trained models for global deployment without adaptation. Our findings suggest that relatively modest local datasets (hundreds to thousands of cases per condition) are sufficient for effective transfer learning when combined with appropriate fine-tuning protocols and fairness constraints.

The federated learning framework, enabling continuous model improvement while preserving privacy, offers a path toward "living" AI systems that adapt to local contexts over time. Quarterly federated updates provided incremental performance gains (1.2–2.1 percentage points) beyond initial fine-tuning, suggesting that sustained improvement is achievable as deployment continues.

However, important limitations remain. Residual bias of 4.7% average accuracy gap, while much improved, still represents differential outcomes between populations. Whether this remaining gap reflects irreducible biological differences, persistent data limitations, or inadequate mitigation techniques requires further investigation.

AI-CDSS Improves Clinical Outcomes Through Multiple Mechanisms

Third, AI assistance translated to meaningful clinical impact through multiple causal pathways. The 23.7% improvement in diagnostic accuracy (p<0.001) represents the proximal effect, but downstream consequences included improved treatment appropriateness (aOR=2.21), reduced misdiagnosis-related harm (56% reduction), and decreased 90-day mortality (aOR=0.64, p=0.004).

The mortality benefit—approximately 121 deaths potentially prevented among 18,947 patients—deserves emphasis. While mortality was a pre-specified outcome, we acknowledge limitations in causal inference from a quasi-experimental design. Nonetheless, the consistency of effects across all safety and outcome measures, combined with biological plausibility (better diagnosis → better treatment → better outcomes), supports a genuine benefit.

Several mechanisms likely contributed to improved outcomes:

1. Cognitive support: AI-CDSS expanded the differential diagnosis consideration, reducing anchoring bias and premature closure ^[30]. Physicians reported that seeing AI-generated differentials prompted consideration of diagnoses they might otherwise have missed.

2. Knowledge augmentation: Embedded clinical guidelines and treatment recommendations provided point-of-care decision support, particularly valuable for less-experienced providers and rare conditions encountered infrequently.

3. Risk stratification: Uncertainty quantification flagged high-risk or diagnostically uncertain cases (17.3% of total) for escalation, ensuring appropriate specialist involvement for complex presentations.

4. Quality improvement: Participating physicians reported educational benefits—learning from AI-generated differentials improved their skills over time (82% agreement), suggesting sustained capacity building beyond immediate decision support.

The reduction in specialist referrals (31.2%) combined with improved referral appropriateness (from 62.4% to 81.8%) demonstrates task-shifting potential. AI-CDSS enabled primary care physicians to confidently manage conditions that previously would have been referred, while simultaneously improving identification of cases genuinely requiring specialist expertise.

Cost-effectiveness analysis revealed substantial economic value: $2 net savings per patient encounter and cost per QALY gained of only $320. These figures compare extremely favorably with other health interventions in low-resource settings ^[31] and suggest AI-CDSS deployment is economically viable even in resource-constrained health systems.

Limitations and Future Directions

Several limitations warrant consideration. First, the quasi-experimental design (AI-assisted vs. historical controls) introduces potential confounding, though we attempted to address this through risk adjustment and sensitivity analyses. Ideally, a randomized controlled trial would provide stronger causal inference.

Second, diagnostic accuracy adjudication relied on expert panel review rather than definitive gold standards (which often don't exist for syndromic presentations). While we employed rigorous adjudication protocols (three blinded specialists per case), measurement error in the outcome is inevitable.

Third, generalizability beyond study sites is uncertain. Participating sites were selected for adequate infrastructure (electricity, security, trained staff), potentially limiting applicability to the most resource-constrained settings. Additionally, all sites received external support for system deployment and physician training; whether health systems can implement AI-CDSS independently remains to be demonstrated.

Fourth, the 24-month follow-up, while longer than most digital health pilots ^[32], is insufficient to assess long-term sustainability, system maintenance, and model degradation as medical knowledge evolves.

Finally, while we attempted to mitigate algorithmic bias, the fundamental challenge remains: LMIC populations are underrepresented in medical datasets globally ^[33], limiting initial model quality before expensive local data collection. More equitable data sharing frameworks, international dataset collaborations, and synthetic data approaches warrant exploration.

Future research should prioritize: (1) randomized controlled trials; (2) implementation studies in even more resource-constrained settings; (3) longitudinal studies beyond 2 years; (4) extension to specialized clinical domains; (5) integration with other digital health tools; (6) human factors research on optimal human-AI collaboration; and (7) policy research on regulatory frameworks and equitable access.

Implications for Global Health Equity

The scarcity of specialist physicians in LMICs represents a fundamental determinant of health inequity, contributing to estimated 8.6 million preventable deaths annually ^[34]. Our findings demonstrate that AI-CDSS offers a scalable intervention to partially address this specialist gap by augmenting primary care provider capabilities. Critically, this intervention proved effective precisely in the settings where it is most needed: rural areas, complex cases, and among less-experienced providers.

However, realizing this potential at scale requires confronting several challenges. Technical feasibility does not guarantee equitable access: commercial AI systems remain expensive, proprietary, and often unavailable to low-resource settings ^[35]. Open-source AI tools, equitable pricing models, and technology transfer initiatives are essential to prevent AI from exacerbating existing health inequalities.

Data colonialism—extracting health data from LMICs to train AI systems that are then sold back at high cost—represents a serious ethical concern ^[36]. Our federated learning approach offers a partial solution by enabling local sites to contribute to and benefit from model improvements while maintaining data ownership.

The algorithmic bias we documented in HIC-trained models represents a form of structural inequity: populations historically excluded from medical research ^[37] now face further disadvantage from AI systems trained on data that excludes them. Our bias mitigation success demonstrates that fairness-aware AI development is technically achievable but requires intentional effort, resources, and institutional commitment.

Ultimately, AI-CDSS should be viewed as complementary to—not a substitute for—fundamental health system strengthening: training more physicians, improving infrastructure, ensuring medication access, and addressing social determinants of health. AI is a tool that can help maximize impact of existing human resources, but cannot replace the need for equitable investment in health systems and workforce development.

Conclusion

This study demonstrates that offline-capable, bias-mitigated AI-driven clinical decision support systems can meaningfully improve diagnostic accuracy, treatment quality, and patient outcomes in resource-constrained primary care settings. Through edge computing architectures, transfer learning with local data, and federated learning protocols, we achieved high performance despite intermittent connectivity and mitigated substantial algorithmic bias from HIC-trained base models. Real-world deployment across diverse settings in four countries showed 23.7% improvement in diagnostic accuracy, 31.2% reduction in specialist referrals, and improved patient safety and outcomes including reduced 90-day mortality.

These findings establish technical feasibility and clinical effectiveness of AI-CDSS in low-resource settings where specialist shortages are most acute. However, realizing the promise of AI for global health equity requires addressing challenges of equitable access, algorithmic fairness, data sovereignty, and integration with broader health system strengthening efforts. With appropriate attention to these considerations, AI-CDSS offers a scalable intervention to augment primary care capabilities and reduce health disparities in settings where advanced medical expertise is scarcest—moving closer to the goal of health for all, regardless of geography or resources.

Data Availability

De-identified patient-level data supporting the findings of this study are available upon reasonable request to the corresponding author, subject to approval by relevant institutional review boards and data use agreements ensuring patient privacy. AI model code and training pipelines are available at https://github.com/global-health-ai/lmic-cdss under open-source MIT license. Interactive tools for exploring bias metrics and performance across subgroups are available at https://lmic-ai-dashboard.global-health.org.

Code Availability

Complete source code for the AI-CDSS, including model architectures, training scripts, bias mitigation algorithms, federated learning implementation, and offline deployment tools, is available at https://github.com/global-health-ai/lmic-cdss under MIT license. Documentation, deployment guides, and training materials are provided in the repository. We encourage adaptation and deployment by other institutions; technical support is available through the project website.

Acknowledgments

We thank the participating healthcare workers, patients, and communities who made this research possible. We acknowledge the data science teams at each site for data collection and quality assurance. We thank the independent Data Safety Monitoring Board (Dr. M. Wilson, Chair; Dr. K. Ouma; Dr. P. Fernandes) for oversight. We acknowledge computing resources provided by Google Cloud for Education and technical support from TensorFlow Lite development team. This work was supported by grants from the Bill & Melinda Gates Foundation (INV-023471), Wellcome Trust (224522/Z/21/Z), and US National Institutes of Health Fogarty International Center (R01TW012184).

Author Contributions

N.R. conceived the study, designed the AI architecture, supervised technical development, conducted statistical analyses, and wrote the manuscript. S.P. led model development, implemented bias mitigation algorithms, and contributed to manuscript writing. M.O. coordinated clinical implementation at Nigerian sites, oversaw data collection, and contributed to study design and manuscript review. L.C. developed offline deployment architecture, implemented federated learning protocols, and contributed to technical methods. A.M. coordinated clinical implementation at Indian sites, conducted physician training, and contributed to clinical interpretation and manuscript review. R.K. coordinated implementation at Kenyan sites, led qualitative research components, and contributed to manuscript review. P.S. coordinated implementation at Brazilian site, conducted cost-effectiveness analysis, and contributed to manuscript review. All authors reviewed and approved the final manuscript.

Competing Interests

The authors declare no competing financial interests. N.R. serves on the scientific advisory board of Digital Health Equity Initiative, a non-profit organization, without financial compensation. S.P. previously received consulting fees from HealthAI Solutions (2022–2023) unrelated to this work.

References

References

1. Organization WH. Global Health Workforce Statistics, 2023 Update. 2023.
2. Liu JX, Goryakin Y, Maeda A, Bruckner T, Scheffler R. Global health workforce labor market projections for 2030. Human Resources for Health. 2017;15(1):11.
3. Organization WH. The World Health Report 2006: Working Together for Health. 2006.
4. Muula AS, Panulo M, Maseko FC. The financial losses from the migration of nurses from Malawi. BMC Nursing. 2006;5:9.
5. Esteva A, Kuprel B, Novoa RA, others. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115-118.
6. McKinney SM, Sieniek M, Godbole V, others. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89-94.
7. Campanella G, Hanna MG, Geneslaw L, others. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature Medicine. 2019;25(8):1301-1309.
8. Liu Y, Jain A, Eng C, others. A deep learning system for differential diagnosis of skin diseases. Nature Medicine. 2020;26(6):900-908.
9. Gulshan V, Peng L, Coram M, others. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402-2410.
10. Sutton RT, Pincock D, Baumgart DC, Sadowski DC, Fedorak RN, Kroeker KI. An overview of clinical decision support systems: benefits, risks, and strategies for success. NPJ Digital Medicine. 2020;3:17.
11. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine. 2019;25(1):44-56.
12. Union IT. Measuring Digital Development: Facts and Figures 2023. 2023.
13. Char DS, Shah NH, Magnus D. Implementing machine learning in health care---addressing ethical challenges. New England Journal of Medicine. 2018;378(11):981-983.
14. Rajkomar A, Hardt M, Howell MD, Corrado G, Chin MH. Ensuring fairness in machine learning to advance health equity. Annals of Internal Medicine. 2018;169(12):866-872.
15. Adamson AS, Smith A. Machine learning and health care disparities in dermatology. JAMA Dermatology. 2018;154(11):1247-1248.
16. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453.
17. Keane PA, Topol EJ. With an eye to AI and autonomous diagnosis. NPJ Digital Medicine. 2018;1:40.
18. Chen T, Guestrin C. XGBoost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016:785-794.
19. Caruana R, Lou Y, Gehrke J, Koch P, Sturm M, Elhadad N. Intelligible models for healthcare: predicting pneumonia risk and hospital 30-day readmission. Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2015:1721-1730.
20. Gal Y, Ghahramani Z. Dropout as a Bayesian approximation: representing model uncertainty in deep learning. Proceedings of the 33rd International Conference on Machine Learning. 2016:1050-1059.
21. Jacob B, Kligys S, Chen B, others. Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018:2704-2713.
22. Zhang BH, Lemoine B, Mitchell M. Mitigating unwanted biases with adversarial learning. Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society. 2018:335-340.
23. McMahan B, Moore E, Ramage D, Hampson S, Arcas BA. Communication-efficient learning of deep networks from decentralized data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics. 2017:1273-1282.
24. Hardt M, Price E, Srebro N. Equality of opportunity in supervised learning. Advances in Neural Information Processing Systems 29. 2016:3315-3323.
25. Wahl B, Cossy-Gantner A, Germann S, Schwalbe NR. Artificial intelligence (AI) and global health: how can AI contribute to health in resource-poor settings?. BMJ Global Health. 2018;3(4):e000798.
26. Jimenez G, Tyagi S, Osman T, others. Improving the primary care consultation for diabetes and depression through digital medical interview assistant systems: narrative review. Journal of Medical Internet Research. 2020;22(8):e18109.
27. Seyyed-Kalantari L, Zhang H, McDermott MBA, Chen IY, Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine. 2021;27(12):2176-2182.
28. Gichoya JW, Banerjee I, Bhimireddy AR, others. AI recognition of patient race in medical imaging: a modelling study. Lancet Digital Health. 2022;4(6):e406-e414.
29. Chen IY, Pierson E, Rose S, Joshi S, Ferryman K, Ghassemi M. Ethical machine learning in healthcare. Annual Review of Biomedical Data Science. 2021;4:123-144.
30. Croskerry P. A universal model of diagnostic reasoning. Academic Medicine. 2009;84(8):1022-1028.
31. Leech AA, Kim DD, Cohen JT, Neumann PJ. Use and misuse of cost-effectiveness analysis thresholds in low- and middle-income countries: trends in cost-per-DALY studies. Value in Health. 2018;21(7):759-761.
32. Labrique AB, Wadhwani C, Williams KA, others. Best practices in scaling digital health in low and middle income countries. Globalization and Health. 2018;14(1):103.
33. Buolamwini J, Gebru T. Gender shades: intersectional accuracy disparities in commercial gender classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency. 2018:77-91.
34. Diseases G, Collaborators I. Global burden of 369 diseases and injuries in 204 countries and territories, 1990--2019: a systematic analysis for the Global Burden of Disease Study 2019. The Lancet. 2020;396(10258):1204-1222.
35. Schwalbe N, Wahl B. Artificial intelligence and the future of global health. The Lancet. 2020;395(10236):1579-1586.
36. Abebe R, Barocas S, Kleinberg J, Levy K, Raghavan M, Robinson DG. Roles for computing in social change. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. 2020:252-260.
37. Hamel LM, Penner LA, Albrecht TL, Heath E, Gwede CK, Eggly S. Barriers to clinical trial enrollment in racial and ethnic minority patients with cancer. Cancer Control. 2016;23(4):327-337.