Development of a Bayesian subjective model for predicting the clinical diagnosis of Ebola in the Democratic Republic of the Congo
John Kamwina Kebela, Prince Kimpanga, Jean Nyandwe, Jack Kokolomami, Steve Bwira, Rostin Mabela, Olivier Mangapi, Berthe Barhayiga, Godfroid Musema, Bienvenu Kabasele, Sylvain Munyanga
Diagnostic Pathology (2026) · School of Public Health, University of Kinshasa, DRC
Received 10 June 2025 · accepted 4 August 2026 · published online 6 August 2026
doi:10.1186/s13000-026-01825-4
BEACON is developed with John Kamwina Kebela, lead and corresponding author of the source study, who owns this platform. This page summarises the study; the article's own text is not reproduced — follow the link to read it in full.
What the model does
Early EVD diagnosis is hard because the first symptoms — fever, vomiting, diarrhoea — look like malaria. Patients are often treated for malaria before EVD is considered, which delays management and exposes health workers.
Seven experts identified seven clusters of clinical and epidemiological signs. A likelihood ratio was derived for each from expert consensus. The model starts from prior odds (PPQ = 2.18), multiplies by each factor's likelihood ratio for present or absent, and converts the posterior odds to a probability.
The external-validation cohort comes from the DRC's 10th outbreak (North Kivu and Ituri, DRC): 695 days, 3,470 cases and 2,280 deaths — a 66% fatality rate.
RV = Π LRi · PoPQ = RV × PPQ · P = PoPQ / (1 + PoPQ)
Decision cutoff: 80% (the paper's optimal cut-off point, Table 6).
The seven factors
Likelihood ratios (Table 3) and regression coefficients (Table 11).
| Factor | LR present | LR absent | β |
|---|---|---|---|
| F1Haemorrhagic signs | 3.438 | 0.278 | -2.07 |
| F2Neurological signs | 1.519 | 0.674 | +0.50 |
| F3Digestive signs | 1.625 | 0.474 | +0.42 |
| F4Pain syndrome | 1.459 | 0.485 | +0.72 |
| F5General signs | 1.450 | 0.600 | +1.37 |
| F6Epidemiological link | 2.913 | 0.064 | -0.62 |
| F7Respiratory signs | 1.440 | 0.756 | -0.21 |
The LR values are recomputed at full precision from the S/N ratios underlying Table 3, not from the paper's pre-rounded two-decimal “LR” column — rounding compounds across seven multiplications and breaks 3 of the 14 Table 4 cases.
Diagnostic performance
The paper's Table 10, across four comparisons.
| Comparison | Se | Sp | PPV | NPV | Acc. |
|---|---|---|---|---|---|
| BSM vs expert consensus42 hypothetical cases, internal validation | 82.6% | 100.0% | 100.0% | 82.6% | 90.5% |
| BSM vs real EVD cases450 patients, DRC — external validation | 82.4% | 33.7% | 80.5% | 36.5% | 71.1% |
| BSM vs LRM on real cases450 patients, DRC | 84.2% | 30.2% | 66.1% | 54.2% | 63.6% |
| LRM vs real EVD cases450 patients, DRC — the comparison model | 71.7% | 71.2% | 89.2% | 43.0% | 71.6% |
Note the gap between the first two rows. On hypothetical cases specificity is 100%; on 450 real patients it falls to 33.7%. That collapse is this model's most important limitation, and the reason this tool surfaces the caveat everywhere.
What that means for 450 patients
External validation at the 80% cutoff (Table 8).
285
Flagged, had EVD
correctly isolated
69
Flagged, did not have EVD
unnecessary isolation
61
Not flagged, had EVD
missed cases
35
Not flagged, no EVD
correctly cleared
Of the 96 patients scored below the cutoff, 61 had EVD — an NPV of 36.5%. That is why a low score is never shown in green anywhere in this tool: it clears nobody.
The logistic regression comparison
Table 11. Only four factors reach significance.
| Factor | B | OR | 95% CI | p |
|---|---|---|---|---|
| Hemorrhagic | -2.07 | 0.13 | 0.07–0.23 | <0.001 |
| Neurological | 0.50 | 1.65 | 0.98–2.78 | 0.06 |
| Digestive | 0.42 | 1.53 | 0.79–2.97 | 0.21 |
| Pain | 0.72 | 2.06 | 1.19–3.58 | 0.01 |
| General | 1.37 | 3.92 | 1.76–8.73 | <0.001 |
| Epidemiological link | -0.62 | 0.54 | 0.30–0.96 | 0.04 |
| Respiratory | -0.21 | 0.81 | 0.46–1.42 | 0.46 |
The Bayesian model is more sensitive (84.2% vs 71.7%), which is the paper's headline conclusion. The regression is markedly more specific, though (71.2% vs 30.2%) — the trade-off is real and runs both ways.
Stated limitations
- —Low real-world specificity (33.7%), producing false positives that burden already-limited resources.
- —External validation was single-centre, in the DRC — generalisability to other epidemiological settings is untested.
- —Expert concordance was only moderate (Cohen's kappa = 0.43); the Delphi technique was used to reduce that variability.
- —The paper recommends multicentre studies and a prospective study during a future outbreak.
- —This is an early-access, unedited manuscript — errors may remain.
How this tool differs
A third state: “not asked”. The published model is binary — every factor is either present or absent. But absence is strong evidence here: F6 absent carries an LR of 0.064 and collapses the posterior. Scoring a factor absent merely because nobody asked produces a confidently wrong low probability. This tool treats “not asked” as LR 1.0, leaving the odds untouched.
Natural frequencies. Results are stated as frequencies (“of 100 patients with a result like this, about 81 had EVD”) rather than bare percentages, which clinicians read less reliably.
Local performance. Once enough RT-PCR outcomes are recorded, the register computes the facility's own Se/Sp/PPV/NPV. Given the gap between 100% and 33.7% specificity, a local number beats a published one.
The implementation reproduces all 14 of the paper's Table 4 cases exactly — probability and RV column — checked automatically across both implementations.