IMPORTANT: Data Provenance Notice

This report is built on a SYNTHETIC dataset and is intended for methodology demonstration only. It must not be used for real clinical decision-making.

Individual-level MalariaGEN severe malaria host genotype and clinical data (the “Genome-wide study of resistance to severe malaria in eleven populations”) requires formal EGA Data Access Committee approval and was not obtainable within this project’s timeframe — appropriate given this is identifiable human genetic and clinical data. The freely available MalariaGEN release for this study contains only aggregate per-SNP summary statistics, with no individual patient rows and no clinical variables (parasitemia, hemoglobin, coma status, etc.).

The dataset used below is a clearly labeled synthetic dataset. Genetic effect sizes are grounded in real, cited published values where available:

Countries modeled as epidemiologically homogeneous: with genuine country-specific grounding available for only Gambia (and partial confirmation for Mali/Kenya), the other 12 West African countries in this dataset share identical true model parameters. Any small country-to-country differences visible in the raw data are sampling noise, not modeled effects, and should not be over-interpreted.

Executive Summary

This report presents a logistic regression model estimating severe malaria risk from host genetic markers (HbS sickle trait, ABO blood group) and routinely collected clinical variables (age, parasitemia, hemoglobin), intended to demonstrate the statistical methodology a real West African triage support tool would require. Trained on 2100 synthetic patients and evaluated on a held-out set of 900 patients never seen during model fitting, the model achieves an AUC of approximately 0.77 (see Results) with good calibration across risk deciles. Sickle trait (HbAS) and blood group O were both directionally protective as expected from the literature, though only HbAS reached statistical significance in this sample. Parasitemia, low hemoglobin, and younger age were the strongest predictors of severe disease, consistent with established clinical knowledge.

Data and Methods

Study Population

3000 synthetic patients across 15 West African countries (Cabo Verde excluded due to distinct, near-eliminated malaria transmission epidemiology). See Data Provenance Notice above.

Predictors

Variable Description Source tier
HbS genotype AA (wild-type) / AS (carrier) / SS (sickle cell disease) Mixed (see notice)
ABO group O Blood group O vs. non-O Representative
Age (years) Patient age at presentation Clinical
log10 parasitemia Presenting parasite density Clinical
Hemoglobin (g/dL) Presenting hemoglobin level Clinical

Model

Multivariable logistic regression, fit using Firth’s penalized likelihood method (logistf package) rather than standard maximum likelihood. This choice was necessary and is reported transparently: an initial standard glm() fit encountered complete separation in the HbSS stratum (zero severe-malaria events among HbSS patients in the training set), producing a numerically meaningless coefficient (unbounded estimate, 95% CI extending to millions). Firth’s method corrects this by construction, without requiring the HbSS category to be collapsed into HbAS or excluded from the model.

The model was trained on a random 70% split of the data (stratified by outcome) and evaluated exclusively on the remaining, untouched 30% test set, following standard practice to avoid overstating performance via in-sample evaluation.

Results

Odds Ratio Table

Firth-corrected odds ratios for severe malaria (training set, n = 2,100)
Predictor Odds Ratio 95% CI Lower 95% CI Upper P value
hbs_genotypeAS HbS genotype: AS (vs. AA) 0.247 0.099 0.517 0.0001
hbs_genotypeSS HbS genotype: SS (vs. AA) 0.392 0.003 3.100 0.4571
abo_group_O ABO group O (vs. non-O) 0.874 0.623 1.222 0.4311
age_years Age (per year) 0.868 0.808 0.929 0.0000
log10_parasitemia Parasitemia (per log10 unit) 2.116 1.762 2.556 0.0000
hemoglobin_gdl Hemoglobin (per g/dL) 0.687 0.632 0.744 0.0000

Interpretation:

  • HbAS carriers had significantly reduced odds of severe malaria (OR ≈ 0.25, p < 0.001), consistent in direction and rough magnitude with the cited Gambia (0.10) and Kenya (0.15) published values.
  • HbSS: not statistically reliable. Among the 20 HbSS patients in the training set, zero experienced severe malaria. While the point estimate suggests a possible protective direction, the sample size is far too small to draw any conclusion specific to HbSS. This model should not be used to inform triage decisions on the basis of HbSS status alone.
  • ABO group O was directionally protective but not statistically significant in this sample (p ≈ 0.43), despite a true protective effect having been built into the underlying simulation — a plausible outcome given ABO’s modest true effect size relative to sample size, not evidence the effect is absent.
  • Parasitemia, hemoglobin, and age were all strong, highly significant predictors in the expected clinical direction.

Discrimination: ROC Curve and AUC

Test-set AUC = 0.766, evaluated on 900 patients not used in model fitting. An AUC in this range is generally considered clinically useful for a triage/risk stratification tool, though real-world validation on genuinely independent clinical data (not simulation) would be required before any deployment.

Calibration

Calibration was strong across all deciles, including the highest-risk group (predicted ≈ 0.30, observed ≈ 0.30). This means the model’s stated probabilities can be trusted as approximately accurate risk estimates, not just useful for ranking patients relative to each other — an important distinction for a clinical tool where a specific percentage might inform a real triage threshold.

Limitations

  1. This entire dataset is synthetic. No real patient data underlies these results. All effect sizes are either directly cited, qualitatively confirmed but not precisely sourced, or representative regional estimates — see the Data Provenance Notice for the exact sourcing tier of each variable.
  2. HbSS estimates are statistically unreliable due to a small number of HbSS patients and zero observed severe events among them in this sample.
  3. Country-level epidemiological differences are not modeled beyond Gambia; 14 of 15 countries share identical true parameters, which does not reflect genuine epidemiological homogeneity across West Africa.
  4. ABO group O’s true effect was not detected as statistically significant in this sample size, despite being built into the simulation — a caution against over-interpreting non-significant results as evidence of no effect.
  5. External validity: even setting aside the synthetic nature of this data, no model — however well-calibrated on its own held-out test set — should be deployed clinically without prospective validation on genuinely independent, real patient data from the target population.

Conclusion

This report demonstrates a complete, methodologically sound pipeline for a malaria severity triage model — including honest handling of a real statistical failure mode (complete separation in the HbSS stratum), held-out test set evaluation rather than in-sample metrics, and transparent labeling of which effect sizes are genuinely literature-grounded versus representative estimates. Real-world deployment of a tool like this would require the actual MalariaGEN individual-level dataset (via EGA Data Access Committee approval) and prospective clinical validation, neither of which substitute for the synthetic exercise presented here.