Aletheia: An Offline-First Clinical Decision Support System for Differential Diagnosis in Low-Resource Healthcare Settings

# An Offline-First Clinical Decision Support System for Differential Diagnosis in Low-Resource Healthcare Settings
Source: [https://arxiv.org/html/2607.24814](https://arxiv.org/html/2607.24814)
Joseph Walusimbi∗, Ann Move Oguti, Abubakhari Sserwadda, Precious Boss Kasasira, and Charles Brian OkoboiJ\. Walusimbi, A\. M\. Oguti, and A\. Sserwadda are with the Department of Electronics and Computer Engineering, Soroti University, Soroti, Uganda\.∗Corresponding author: J\. Walusimbi \(e\-mail: 2401600068@sun\.ac\.ug\)\. A\. M\. Oguti \(e\-mail: amoguti@sun\.ac\.ug\)\. A\. Sserwadda \(e\-mail: asserwadda@sun\.ac\.ug\)\.P\. B\. Kasasira and C\. B\. Okoboi are with the School of Health Sciences, Soroti University, Soroti, Uganda \(e\-mail: kasasirakp40@gmail\.com; charlesbryahn@gmail\.com\)\.

###### Abstract

Access to specialist clinical expertise remains severely limited across sub\-Saharan Africa, where physician\-to\-patient ratios can fall below 1:25,000 in rural settings\. Existing AI\-assisted diagnostic tools predominantly require reliable internet connectivity and high\-specification hardware, rendering them impractical for frontline healthcare workers in district hospitals and health centres\. This paper presentsAletheia, an offline\-first clinical decision support system designed for low\-resource healthcare contexts across sub\-Saharan Africa\. Aletheia is built upon Qwen2\.5\-3B\-Instruct, fine\-tuned using Quantised Low\-Rank Adaptation \(QLoRA\) on a curated dataset of 27,000 clinical reasoning samples spanning 50 disease conditions with elevated prevalence in East Africa\. Evaluation demonstrates a Top\-1 diagnostic accuracy of 80\.0%, Top\-3 accuracy of 100\.0%, BERTScore\-F1 of 0\.909, and METEOR of 0\.467 across ten representative clinical case categories\. The system achieves an Expected Calibration Error \(ECE\) of 0\.275 and passes the Africa Deep Tech Challenge 2026 \(ADTC 2026\) memory budget constraint of 7 168 MB, achieving a peak inference RAM of approximately 3 630 MB on the standardised benchmark laptop\. These results demonstrate the feasibility of deploying large language model\-based clinical reasoning at the primary care level in resource\-constrained settings without cloud infrastructure\.

## IIntroduction

The burden of disease in sub\-Saharan Africa is disproportionately high relative to the availability of clinical expertise\. Uganda, with a population exceeding 48 million, has a physician density of approximately 0\.17 per 1,000 population\[[1](https://arxiv.org/html/2607.24814#bib.bib1)\]—among the lowest globally\. In rural and peri\-urban settings such as Soroti District in Eastern Uganda, a single clinical officer may be responsible for 80–120 patient consultations per day, leaving fewer than five minutes per patient for history\-taking, examination, differential diagnosis, and investigation planning\. Under these conditions, diagnostic errors and missed critical presentations are an inevitable consequence of cognitive overload rather than clinical incompetence\.

Artificial intelligence\-assisted clinical decision support systems \(CDSS\) have demonstrated considerable promise in high\-income settings, with models achieving diagnostic accuracy comparable to specialist physicians in radiology\[[2](https://arxiv.org/html/2607.24814#bib.bib2)\], dermatology\[[3](https://arxiv.org/html/2607.24814#bib.bib3)\], and general internal medicine\[[4](https://arxiv.org/html/2607.24814#bib.bib4)\]\. However, the vast majority of such systems require persistent internet connectivity, cloud\-based inference infrastructure, and hardware specifications far beyond what is available in sub\-Saharan African health facilities\[[5](https://arxiv.org/html/2607.24814#bib.bib5)\]\.

Recent advances in large language model \(LLM\) compression, particularly 4\-bit quantisation via the GGUF format and low\-rank adaptation \(LoRA\) fine\-tuning, have created a new possibility: deploying clinically useful AI reasoning on commodity hardware without internet connectivity\[[6](https://arxiv.org/html/2607.24814#bib.bib6),[7](https://arxiv.org/html/2607.24814#bib.bib7)\]\. The combination of these techniques enables a 3\-billion parameter language model to be compressed to under 2 GB and executed on a standard laptop at practical inference speeds\.

This paper presentsAletheia\(from the Greek fortruthordisclosure\), an offline\-first clinical decision support system that:

- •Runs entirely on\-device with no internet dependency, suitable for deployment at district hospital and health centre level across sub\-Saharan Africa;
- •Provides ranked differential diagnoses with probability estimates, evidence\-based investigation recommendations, clinical rationale, red flag identification, and follow\-up question generation;
- •Is fine\-tuned on a dataset weighted toward African disease epidemiology — with particular emphasis on East African prevalence patterns — including conditions rarely represented in publicly available medical AI benchmarks;
- •Satisfies the memory constraints of the Africa Deep Tech Challenge 2026 \(ADTC 2026\)\[[29](https://arxiv.org/html/2607.24814#bib.bib29)\], which requires models to operate within 7 168 MB on a standardised laptop \(Intel Core i5 10th–12th generation, 8 GB DDR4, Ubuntu 22\.04, no discrete GPU\), with Aletheia achieving a peak RAM requirement of approximately 3 730 MB — a margin of 3 438 MB below the ceiling\.

The remainder of this paper is structured as follows\. Section[II](https://arxiv.org/html/2607.24814#S2)reviews related work in medical LLMs and clinical decision support for low\-resource settings\. Section[III](https://arxiv.org/html/2607.24814#S3)describes the training dataset\. Section[IV](https://arxiv.org/html/2607.24814#S4)presents the model architecture, fine\-tuning methodology, and deployment pipeline\. Section[V](https://arxiv.org/html/2607.24814#S5)details the experimental setup and evaluation metrics\. Section[VI](https://arxiv.org/html/2607.24814#S6)reports quantitative results\. Section[VII](https://arxiv.org/html/2607.24814#S7)discusses findings, limitations, and future directions\. Section[VIII](https://arxiv.org/html/2607.24814#S8)concludes the paper\.

## IIRelated Work

### II\-ALarge Language Models in Clinical Medicine

The application of transformer\-based language models to clinical reasoning has accelerated substantially since the release of GPT\-4\[[8](https://arxiv.org/html/2607.24814#bib.bib8)\]and its successors\. Med\-PaLM 2\[[9](https://arxiv.org/html/2607.24814#bib.bib9)\]achieved expert\-level performance on USMLE\-style medical questions, while BioMedLM\[[10](https://arxiv.org/html/2607.24814#bib.bib10)\]and ClinicalBERT\[[11](https://arxiv.org/html/2607.24814#bib.bib11)\]demonstrated the value of domain\-specific pre\-training\. However, these systems are predominantly evaluated on US or European clinical benchmarks and require cloud\-based inference infrastructure, limiting their applicability in resource\-constrained settings\.

### II\-BClinical Decision Support in Low\-Resource Settings

Prior work on CDSS for sub\-Saharan Africa has focused primarily on rule\-based expert systems\[[12](https://arxiv.org/html/2607.24814#bib.bib12)\], mobile health \(mHealth\) applications\[[13](https://arxiv.org/html/2607.24814#bib.bib13)\], and telemedicine platforms\[[14](https://arxiv.org/html/2607.24814#bib.bib14)\]\. While these approaches have demonstrated utility, they either lack the reasoning depth of modern LLMs or depend on internet connectivity for core functionality\. Symptom\-checker applications such as Ada\[[15](https://arxiv.org/html/2607.24814#bib.bib15)\]and Babylon\[[16](https://arxiv.org/html/2607.24814#bib.bib16)\]provide structured differential diagnosis but require continuous data transmission to cloud inference endpoints\.

### II\-CEfficient LLM Deployment

The quantisation of large language models for edge deployment has been advanced by work on GPTQ\[[17](https://arxiv.org/html/2607.24814#bib.bib17)\], AWQ\[[18](https://arxiv.org/html/2607.24814#bib.bib18)\], and the GGUF format supported by llama\.cpp\[[7](https://arxiv.org/html/2607.24814#bib.bib7)\]\. QLoRA\[[6](https://arxiv.org/html/2607.24814#bib.bib6)\]demonstrated that high\-quality fine\-tuning of quantised models is achievable with minimal hardware, enabling fine\-tuning of 7B\+ parameter models on consumer GPUs\. The combination of QLoRA fine\-tuning and GGUF deployment enables a workflow where training occurs on available cloud hardware and the resulting model is deployed on commodity local hardware\.

### II\-DAfrican Medical AI

Dedicated African medical AI datasets and benchmarks remain scarce\. WHO AFRO health statistics\[[19](https://arxiv.org/html/2607.24814#bib.bib19)\]and regional epidemiological reports provide limited structured clinical AI benchmark data\. The MedQA\[[20](https://arxiv.org/html/2607.24814#bib.bib20)\]and MedMCQA\[[21](https://arxiv.org/html/2607.24814#bib.bib21)\]datasets offer large\-scale medical question\-answering data but are predominantly US and Indian in clinical context respectively\. Aletheia addresses this gap through a purpose\-built synthetic dataset weighted to reflect East African disease epidemiology\.

## IIIDataset

### III\-ADataset Composition

The Aletheia training dataset was constructed from three sources:

1. 1\.Aletheia\-Synthetic: 20,000 structured clinical reasoning samples generated from 50 hand\-crafted clinical case templates covering conditions prevalent in East Africa\.
2. 2\.MedQA\-USMLE\[[20](https://arxiv.org/html/2607.24814#bib.bib20)\]: 9,358 filtered questions from the MedQA dataset, selected for relevance to tropical medicine, infectious disease, obstetrics, and paediatrics using a 180\-term keyword filter\.
3. 3\.MedMCQA\[[21](https://arxiv.org/html/2607.24814#bib.bib21)\]: 10,000 filtered questions from the MedMCQA dataset, selected using the same keyword filter and capped at 10,000 samples\.

After applying a 60/20/20 mixing ratio across all three sources and a minimum\-token quality filter \(50 tokens\), the final dataset comprised 27,000 training samples and 3,000 evaluation samples, totalling 30,000 samples — an increase of 25% over the initial single\-source configuration\. Table[I](https://arxiv.org/html/2607.24814#S3.T1)summarises the dataset statistics\.

TABLE I:Training Dataset Statistics
### III\-BClinical Conditions

The 50 clinical conditions were selected to reflect the East African disease burden across eight categories: infectious and tropical disease \(12 conditions\), respiratory \(3\), cardiovascular \(3\), obstetric and gynaecological \(4\), paediatric \(4\), neurological \(2\), renal and endocrine \(4\), surgical and trauma \(3\), and other specialties \(15\)\. Conditions were weighted by estimated incidence in East Africa \(africa\_weight parameter, range 3–5\), with malaria, HIV/AIDS, tuberculosis, cerebral malaria, eclampsia, severe acute malnutrition, and neonatal sepsis receiving the highest weights\.

### III\-CReasoning Task Types

Each clinical case was instantiated across eight reasoning task types to teach multi\-step clinical reasoning rather than simple diagnosis recall:

1. 1\.Initial differential diagnosis— ranked differentials from presenting symptoms;
2. 2\.Test recommendation— evidence\-based investigation prioritisation;
3. 3\.Evidence update— Bayesian probability revision after test results;
4. 4\.Rationale explanation— plain\-language clinical reasoning;
5. 5\.Follow\-up questions— most discriminating next question generation;
6. 6\.Severity assessment— acuity classification and level\-of\-care determination;
7. 7\.Treatment hint— district\-hospital\-level immediate management;
8. 8\.Red flag identification— immediate escalation triggers\.

## IVMethodology

### IV\-ABase Model Selection

Qwen2\.5\-3B\-Instruct\[[22](https://arxiv.org/html/2607.24814#bib.bib22)\]was selected as the base model for the following reasons:

- •Parameter count \(3\.09B\) is sufficient for multi\-step clinical reasoning while remaining within the ADTC memory budget after quantisation;
- •Native bfloat16 support enables stable training on A100 hardware without numerical instability;
- •Strong baseline performance on instruction\-following tasks reduces the fine\-tuning sample requirement;
- •Permissive licensing supports open deployment in public health contexts\.

### IV\-BFine\-Tuning with QLoRA

We applied Quantised Low\-Rank Adaptation \(QLoRA\)\[[6](https://arxiv.org/html/2607.24814#bib.bib6)\]to fine\-tune Qwen2\.5\-3B\-Instruct\. On the A100 training hardware, we used full bfloat16 precision without quantisation to maximise training quality\. The LoRA adapters were attached to all linear projection layers:q\_proj,k\_proj,v\_proj,o\_proj,gate\_proj,up\_proj, anddown\_proj\.

Table[II](https://arxiv.org/html/2607.24814#S4.T2)summarises the training configuration\.

TABLE II:Experimental Configuration
### IV\-CDeployment Pipeline

After fine\-tuning, the LoRA adapters were merged into the base model weights using the PEFT library\[[23](https://arxiv.org/html/2607.24814#bib.bib23)\]\. The merged model was then converted to the GGUF format using llama\.cpp\[[7](https://arxiv.org/html/2607.24814#bib.bib7)\]via a two\-step process: conversion to 16\-bit floating point GGUF \(6\.18 GB\), followed by quantisation to Q4\_K\_M \(1\.93 GB\) and Q2\_K \(1\.27 GB\) usingllama\-quantize\. The primary deployment target is the Q4\_K\_M quantisation, which achieves approximately 98% of F16 quality at 31% of the file size\.

The system provides two interfaces for clinical use: a terminal CLI for scripting and integration, and a web\-based graphical user interface \(GUI\) built with Gradio that runs locally in the browser without internet connectivity\. Both interfaces accept structured clinical input and return structured JSON containing ranked differentials, recommended tests, rationale, and red flags\.

### IV\-DOffline\-First Architecture

Aletheia implements a local\-first, sync\-when\-available architecture:

- •Offline mode: Full diagnostic capability, no degradation\. The 1\.93 GB GGUF model and inference engine operate entirely on\-device;
- •Connected mode: Optional model update download, anonymised aggregate usage logging for quality improvement;
- •Update safety: New model versions are validated on a held\-out test set before activation; rollback to the previous version is retained for 30 days\.

## VExperimental Setup

### V\-ATraining Infrastructure

Training was performed on Google Colab Pro using an NVIDIA A100\-SXM4\-80GB GPU \(85\.1 GB VRAM\) with PyTorch 2\.11\.0, Transformers 4\.44\.2, PEFT 0\.12\.0, TRL 0\.10\.1, and Accelerate 0\.34\.2\. The complete training run required 1\.92 hours for 4,050 steps across 3 epochs\.

### V\-BEvaluation Protocol

Model performance was evaluated across four dimensions:

#### V\-B1Clinical Accuracy

Top\-kkaccuracy was computed by checking whether the correct diagnosis appeared in the model’s top\-1 or top\-3 ranked differentials\. Per\-condition precision, recall, and F1 were computed over the 10 core evaluation case categories\. Accuracy was further stratified by clinical severity \(Critical, High, Moderate, Low\) and reasoning task type\.

#### V\-B2Language Quality

Text generation quality was assessed using ROUGE\-1, ROUGE\-2, and ROUGE\-L\[[24](https://arxiv.org/html/2607.24814#bib.bib24)\]\(F1 scores\), BERTScore\-F1\[[25](https://arxiv.org/html/2607.24814#bib.bib25)\]using theroberta\-largeencoder, and METEOR\[[26](https://arxiv.org/html/2607.24814#bib.bib26)\]\.

#### V\-B3Calibration

Probability calibration was assessed using Expected Calibration Error \(ECE\)\[[27](https://arxiv.org/html/2607.24814#bib.bib27)\], Maximum Calibration Error \(MCE\), and Brier score\[[28](https://arxiv.org/html/2607.24814#bib.bib28)\]stratified by clinical severity\. A reliability diagram was constructed using 10 equal\-width bins\.

#### V\-B4Baseline Comparison

The fine\-tuned Aletheia model was compared against the unmodified Qwen2\.5\-3B\-Instruct base model \(zero\-shot\) on Top\-1 accuracy and ROUGE\-1 to quantify the contribution of fine\-tuning\.

### V\-CHardware Compliance Testing

Inference memory requirements were estimated against the ADTC 2026 standard laptop specification \(Intel Core i5 10th–12th generation, 8 GB DDR4, Ubuntu 22\.04, no discrete GPU\) using the formula:

Mtotal=MOS\+Mmodel\+MKV\+Mruntime\+MappM\_\{\\text\{total\}\}=M\_\{\\text\{OS\}\}\+M\_\{\\text\{model\}\}\+M\_\{\\text\{KV\}\}\+M\_\{\\text\{runtime\}\}\+M\_\{\\text\{app\}\}\(1\)
whereMOS=900M\_\{\\text\{OS\}\}=900MB,MKV=400M\_\{\\text\{KV\}\}=400MB \(1,024 token context\),Mruntime=300M\_\{\\text\{runtime\}\}=300MB, andMapp=200M\_\{\\text\{app\}\}=200MB\.

## VIResults

### VI\-ATraining Dynamics

Fig\.[1](https://arxiv.org/html/2607.24814#S6.F1)shows the training and validation loss curves, perplexity, learning rate schedule, and gradient norm over 4,050 training steps\. Training loss decreased from an initial value of approximately 1\.33 in early steps to a final value of 0\.5197\. The relatively higher terminal loss compared to single\-source training runs reflects the increased task diversity introduced by the MedMCQA component, which exposes the model to a broader range of clinical question styles and reasoning patterns\. Crucially, this broader training signal produced substantial gains in diagnostic accuracy, with Top\-1 accuracy improving from 70\.0% to 80\.0% and Top\-3 accuracy reaching 100\.0% — demonstrating that loss alone is an incomplete proxy for clinical utility\. The cosine learning rate schedule with 5% warmup produced smooth convergence without instability\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig1_training_curves.png)Figure 1:Training dynamics over 4,050 steps \(3 epochs\)\. Top\-left: training and validation loss\. Top\-right: perplexity\. Bottom\-left: cosine learning rate schedule\. Bottom\-right: gradient norm with smoothed overlay\.![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig2_epoch_loss.png)Figure 2:Per\-epoch training loss and validation loss\. Epoch 1: 0\.3498 / 0\.369\. Epoch 2: 0\.3224 / 0\.369\. Epoch 3: 0\.5197\.
### VI\-BClinical Accuracy

Table[III](https://arxiv.org/html/2607.24814#S6.T3)summarises the primary evaluation results\. Aletheia achieved a Top\-1 diagnostic accuracy of 80\.0% and Top\-3 accuracy of 100\.0%, indicating that the correct diagnosis appears within the model’s three highest\-ranked suggestions in all ten clinical presentations — a perfect Top\-3 score\.

TABLE III:Main Evaluation Results![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig02_topk_accuracy.png)Figure 3:Top\-1 \(80\.0%\) and Top\-3 \(100\.0%\) diagnostic accuracy across 10 clinical case categories\.![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig01_per_condition_f1.png)Figure 4:Per\-condition precision, F1, and recall across all 10 evaluation case categories\.![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig03_confusion_matrix.png)Figure 5:Confusion matrix for Top\-1 differential diagnosis predictions across 10 clinical conditions\.
### VI\-CSeverity\-Stratified Performance

Fig\.[6](https://arxiv.org/html/2607.24814#S6.F6)shows diagnostic accuracy stratified by clinical severity\. The model demonstrates its highest accuracy on conditions classified as High and Moderate severity, reflecting the higher representation of these conditions in the training dataset\. Performance on Critical conditions, while lower in absolute Top\-1 accuracy, benefits from the high Top\-3 accuracy of 100\.0%, meaning the correct diagnosis is always present in the ranked output for clinician review\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig04_severity_accuracy.png)Figure 6:Top\-1 diagnostic accuracy stratified by clinical severity \(Critical, High, Moderate, Low\)\.
### VI\-DReasoning Task Performance

Fig\.[7](https://arxiv.org/html/2607.24814#S6.F7)shows Top\-1 accuracy across the five evaluated reasoning task types\. The model performs most strongly on test recommendation and evidence update tasks, reflecting the structured nature of these outputs\. Initial differential diagnosis and follow\-up question generation show lower but clinically meaningful accuracy\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig05_reasoning_type.png)Figure 7:Top\-1 accuracy across five clinical reasoning task types\.
### VI\-ELanguage Quality

Fig\.[8](https://arxiv.org/html/2607.24814#S6.F8)shows the language quality metrics\. BERTScore\-F1 of 0\.909 indicates high semantic similarity between model outputs and reference clinical answers\. The relatively lower ROUGE\-1 score of 0\.383 is expected for generative clinical reasoning tasks where exact lexical overlap is a less meaningful measure than semantic fidelity\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig06_language_quality.png)Figure 8:Language quality metrics: ROUGE\-1, ROUGE\-2, ROUGE\-L, BERTScore\-F1, and METEOR\.
### VI\-FCalibration

Fig\.[9](https://arxiv.org/html/2607.24814#S6.F9)shows the reliability diagram and Brier scores by severity\. The ECE of 0\.275 indicates moderate calibration — the model tends to assign probability estimates that are directionally correct but somewhat overconfident on conditions where it performs well\. Calibration is worst for Critical\-severity conditions, which is the expected pattern for models trained predominantly on synthetic data without clinician\-verified probability distributions\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig07_calibration.png)Figure 9:Left: Reliability diagram with ECE = 0\.275\. Right: Brier score stratified by clinical severity\.
### VI\-GBaseline Comparison

Fig\.[10](https://arxiv.org/html/2607.24814#S6.F10)compares fine\-tuned Aletheia against the unmodified Qwen2\.5\-3B\-Instruct base model \(zero\-shot\)\. Fine\-tuning produces a substantial improvement in Top\-1 diagnostic accuracy and ROUGE\-1, confirming that the clinical reasoning capabilities of Aletheia are a direct product of the fine\-tuning process rather than pre\-existing in the base model\.

![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig08_baseline_comparison.png)Figure 10:Comparison of fine\-tuned Aletheia vs base Qwen2\.5\-3B\-Instruct \(zero\-shot\) on Top\-1 accuracy and ROUGE\-1\.![Refer to caption](https://arxiv.org/html/2607.24814v1/figures/fig09_rouge_distribution.png)Figure 11:Box plot distributions of ROUGE\-1, ROUGE\-2, and ROUGE\-L scores across all evaluation cases, illustrating score spread and consistency\.
### VI\-HHardware Compliance

Table[IV](https://arxiv.org/html/2607.24814#S6.T4)reports the deployment efficiency metrics\. Both GGUF quantisation formats pass the ADTC 2026 memory ceiling of 7,168 MB with substantial margin\.

TABLE IV:Deployment Efficiency \(ADTC 2026 Compliance\)

## VIIDiscussion

### VII\-AClinical Relevance of Results

A Top\-3 accuracy of 100\.0% is the most clinically meaningful result from this evaluation\. In practice, Aletheia presents the clinician with a ranked list of differential diagnoses rather than a single answer\. A clinical officer reviewing a ranked list of three diagnoses and finding the correct one present in 90% of cases represents a genuine decision\-support capability — the system surfaces the diagnostic space that the clinician should be considering, reducing the cognitive load of generating a differential from scratch under time pressure\.

The BERTScore\-F1 of 0\.909 indicates that the model’s clinical reasoning text is semantically faithful to expert reference answers\. This is important beyond accuracy scores: a system that names the correct diagnosis but provides incorrect reasoning is clinically dangerous\. Aletheia’s high BERTScore suggests that its explanations are not only correct in conclusion but coherent in reasoning\.

### VII\-BCalibration and Clinical Safety

The ECE of 0\.275 indicates that the model’s probability estimates require careful interpretation\. In clinical practice, the probability values output by Aletheia should be treated asrelative rankingsrather than absolute probability estimates\. We recommend that the clinical interface present these as “more likely” / “less likely” qualitative rankings rather than specific percentages, until calibration can be improved through clinician\-validated feedback data\.

### VII\-CComparison to Related Work

Direct comparison to prior work is constrained by the lack of standardised African clinical benchmarks\. Med\-PaLM 2\[[9](https://arxiv.org/html/2607.24814#bib.bib9)\]achieved 86\.5% on USMLE questions but requires cloud inference and has no offline deployment pathway\. On the MedQA benchmark, Aletheia’s base model \(Qwen2\.5\-3B\) achieves competitive performance for its parameter count, and fine\-tuning produces substantial gains on Africa\-specific clinical scenarios that are not represented in standard benchmarks\.

The key differentiator of Aletheia is not raw accuracy butdeployability: the ability to run on a 8 GB laptop with no internet, consuming under 4 GB RAM, at a model size of under 2 GB\. No prior medical LLM system has demonstrated this combination of capabilities in resource\-constrained African healthcare contexts — a gap that is relevant across the continent\.

### VII\-DLimitations

Several limitations of this work should be acknowledged:

1. 1\.Synthetic training data: The majority of training samples are synthetically generated from hand\-crafted case templates\. While carefully designed to reflect African clinical presentation patterns, the probability distributions assigned to differentials reflect the authors’ clinical knowledge synthesis rather than empirical epidemiological data derived from prospective patient cohorts\. Ongoing clinician involvement from co\-authors P\. B\. Kasasira and C\. B\. Okoboi is actively informing refinements to the case templates and probability estimates in preparation for the next training iteration\.
2. 2\.Evaluation case set size: The quantitative evaluation reported in this paper was conducted on 10 core case categories using automated metrics\. A broader and clinician\-graded evaluation set is currently under development as part of the ongoing validation activity with the School of Health Sciences co\-authors, and will be reported in a subsequent study\.
3. 3\.Calibration: An ECE of 0\.275 indicates that the model’s probability estimates require improvement before they can be presented as clinically actionable confidence values\. Initial feedback from clinician co\-authors suggests that the directional ranking of differentials is clinically plausible in the majority of cases evaluated, though exact probability estimates should be treated as relative rankings rather than absolute likelihoods\. Calibration improvement is a priority for the next training cycle\.
4. 4\.Clinical validation scope: An initial clinical evaluation of Aletheia is currently underway with two co\-authors from the School of Health Sciences, Soroti University — P\. B\. Kasasira and C\. B\. Okoboi — who are practising clinicians actively testing and evaluating the system against real clinical presentations encountered in their practice\. Early findings are encouraging: the ranked differential output has been rated as clinically plausible and useful for decision support in the majority of cases reviewed to date\. However, this initial evaluation is limited in scale and has not yet been conducted under a formal study protocol with a defined patient cohort\. A larger multi\-site validation study is planned as the immediate next phase, involving clinical officers and physicians across district hospitals, regional referral hospitals, and primary health centres in Eastern Uganda, with a target enrolment of 500\+ patient cases\. This study will assess diagnostic concordance, clinical usability, acceptance, and safety, and is subject to IRB approval from Soroti University and the Uganda National Council for Science and Technology \(UNCST\)\.

### VII\-EFuture Work

Building on the current state of development and the ongoing clinical evaluation, the following directions are planned:

- •Completion and formal reporting of the ongoing clinical evaluation with P\. B\. Kasasira and C\. B\. Okoboi, including a structured assessment of diagnostic concordance, usability, and clinician acceptance across a defined case set;
- •Expansion to a multi\-site prospective validation study across district hospitals, regional referral hospitals, and primary health centres in Eastern Uganda, targeting 500\+ patient cases and 50\+ clinical officers, subject to IRB approval from Soroti University and the Uganda National Council for Science and Technology \(UNCST\);
- •Incorporation of clinician feedback from the ongoing evaluation into the next training iteration, with particular focus on improving calibration \(ECE\), refining probability estimates, and expanding condition coverage from 50 to 100\+ clinical conditions;
- •Development of a lightweight desktop application to complement the existing Gradio web UI and terminal CLI, making Aletheia accessible to non\-technical clinical users;
- •Kiswahili and Ateso language support to serve clinical officers across Eastern Uganda and the wider East African region;
- •Submission to the Uganda National Drug Authority \(NDA\) under the software\-as\-a\-medical\-device regulatory pathway;
- •Commercialisation through Arapai Technologies International Limited, with a distribution model targeting district health facilities across Uganda and the wider East African region at sustainable cost\.

## VIIIConclusion

This paper has presented Aletheia, an offline\-first clinical decision support system designed for frontline healthcare workers in sub\-Saharan Africa\. By combining QLoRA fine\-tuning of a 3B parameter language model on an Africa\-weighted clinical reasoning dataset \(with particular emphasis on East African disease burden\) with GGUF quantisation for edge deployment, Aletheia achieves a Top\-3 diagnostic accuracy of 100\.0% and BERTScore\-F1 of 0\.909 while operating entirely within the memory constraints of a standard 8 GB laptop without internet connectivity\.

The results demonstrate that clinically meaningful AI\-assisted diagnostic reasoning is achievable at the primary care level in resource\-constrained settings without cloud infrastructure\. In the context of Uganda’s physician\-to\-patient ratio of 1:25,000, a system that presents the correct diagnosis in its top suggestion in 80% of cases, and within its top three suggestions in 100% of cases, — running on hardware already present in most health facilities — represents a meaningful force multiplier for the clinical workforce\. The authors invite collaboration from Ugandan clinicians, health informaticists, and the Ministry of Health for the planned prospective validation study\.

## Acknowledgment

The authors thank the clinical staff of Soroti Regional Referral Hospital, Soroti District, whose daily practice under resource constraints shed light on the problem this work addresses\. The authors acknowledge the Africa Deep Tech Challenge 2026 \(ADTC 2026\) for establishing the on\-device AI benchmark standard that shaped the hardware compliance requirements of this work, and for providing a clear deployment target that grounds this research in real\-world constraints relevant to the African technology ecosystem\.

## Funding

No external funding was received for this work\.

## References

- \[1\]World Health Organization, “Global Health Workforce Statistics,” Geneva: WHO, 2023\. \[Online\]\. Available:https://www\.who\.int/data/gho/data/themes/topics/health\-workforce
- \[2\]P\. Rajpurkar, E\. Chen, O\. Banerjee, and E\. J\. Topol, “AI in health and medicine,”Nature Medicine, vol\. 28, no\. 1, pp\. 31–38, 2022\.
- \[3\]A\. Estevaet al\., “Dermatologist\-level classification of skin cancer with deep neural networks,”Nature, vol\. 542, no\. 7639, pp\. 115–118, 2017\.
- \[4\]M\. B\. A\. McDermottet al\., “A comprehensive EHR timeseries pre\-training benchmark,” inProc\. ACM Conf\. Health, Inference, and Learning, 2021, pp\. 257–278\.
- \[5\]World Health Organization, “Global Strategy on Digital Health 2020–2025,” Geneva: WHO, 2021\. \[Online\]\. Available:https://www\.who\.int/docs/default\-source/documents/gs4dhdaa2a9f352b0445bafbc79ca799dce4d\.pdf
- \[6\]T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. Zettlemoyer, “QLoRA: Efficient finetuning of quantized LLMs,” inAdvances in Neural Information Processing Systems, vol\. 36, 2023\.
- \[7\]G\. Gerganov, “llama\.cpp: Inference of LLaMA model in pure C/C\+\+,” GitHub, 2023\. \[Online\]\. Available:https://github\.com/ggerganov/llama\.cpp
- \[8\]OpenAI, “GPT\-4 Technical Report,” arXiv:2303\.08774, 2023\.
- \[9\]K\. Singhalet al\., “Towards expert\-level medical question answering with large language models,” arXiv:2305\.09617, 2023\.
- \[10\]E\. Boltonet al\., “BioMedLM: A domain\-specific large language model for biomedical text,” Stanford CRFM, 2022\.
- \[11\]K\. Huang, J\. Altosaar, and R\. Ranganath, “ClinicalBERT: Modeling clinical notes and predicting hospital readmission,” arXiv:1904\.05342, 2019\.
- \[12\]H\. S\. F\. Fraser, P\. Biondich, D\. Moodley, S\. Choi, B\. W\. Mamlin, and P\. Szolovits, “Implementing electronic medical record systems in developing countries,”Informatics in Primary Care, vol\. 13, no\. 2, pp\. 83–95, 2005\.
- \[13\]A\. B\. Labriqueet al\., “mHealth innovations as health system strengthening tools,”Global Health: Science and Practice, vol\. 1, no\. 2, pp\. 160–171, 2013\.
- \[14\]A\. Bastawrous and M\. J\. Armstrong, “Mobile health use in low\-and high\-income countries,”Journal of the Royal Society of Medicine, vol\. 106, no\. 4, pp\. 130–142, 2013\.
- \[15\]Ada Health GmbH, “Ada: Your personal health guide,” 2020\. \[Online\]\. Available:https://ada\.com
- \[16\]Babylon Health, “Babylon health AI,” 2020\. \[Online\]\. Available:https://www\.babylonhealth\.com
- \[17\]E\. Frantar, S\. Ashkboos, T\. Hoefler, and D\. Alistarh, “GPTQ: Accurate post\-training quantization for generative pre\-trained transformers,” arXiv:2210\.17323, 2022\.
- \[18\]J\. Linet al\., “AWQ: Activation\-aware weight quantization for LLM compression and acceleration,” arXiv:2306\.00978, 2023\.
- \[19\]World Health Organization Regional Office for Africa, “Health Statistics and Information Systems: African Region,” Brazzaville: WHO AFRO, 2023\. \[Online\]\. Available:https://www\.afro\.who\.int/health\-topics/health\-statistics
- \[20\]D\. Jinet al\., “What disease does this patient have? A large\-scale open domain question answering dataset from medical exams,”Applied Sciences, vol\. 11, no\. 14, p\. 6421, 2021\.
- \[21\]A\. Palet al\., “MedMCQA: A large\-scale multi\-subject multi\-choice dataset for medical domain question answering,” inProc\. Conference on Health, Inference, and Learning, PMLR, vol\. 174, pp\. 248–260, 2022\.
- \[22\]Qwen Team, “Qwen2\.5 Technical Report,” arXiv:2412\.15115, 2024\.
- \[23\]S\. Mangrulkaret al\., “PEFT: State\-of\-the\-art parameter\-efficient fine\-tuning methods,” GitHub, 2022\. \[Online\]\. Available:https://github\.com/huggingface/peft
- \[24\]C\.\-Y\. Lin, “ROUGE: A package for automatic evaluation of summaries,” inText Summarization Branches Out, 2004, pp\. 74–81\.
- \[25\]T\. Zhanget al\., “BERTScore: Evaluating text generation with BERT,” inProc\. ICLR, 2020\.
- \[26\]S\. Banerjee and A\. Lavie, “METEOR: An automatic metric for MT evaluation with improved correlation with human judgments,” inProc\. ACL Workshop on Intrinsic and Extrinsic Evaluation Measures, 2005\.
- \[27\]M\. P\. Naeini, G\. F\. Cooper, and M\. Hauskrecht, “Obtaining well calibrated probabilities using Bayesian binning,” inProc\. AAAI, 2015, pp\. 2901–2907\.
- \[28\]G\. W\. Brier, “Verification of forecasts expressed in terms of probability,”Monthly Weather Review, vol\. 78, no\. 1, pp\. 1–3, 1950\.
- \[29\]Africa Deep Tech Challenge, “ADTC 2026: On\-Device Language Model Benchmark for African Hardware Contexts,” 2026\. \[Online\]\. Available:https://africadeeptech\.org/challenge\-2026