We conducted a retrospective cohort study using hospital data obtained from patients admitted to RDH, in the northern region of Malawi between January 2012 and December 2024. The patients were followed up on Day 5 and Day 11 of hospitalisation.

Rumphi District Hospital is a public, secondary-level facility in northern Malawi serving a predominantly rural population. The hospital provides inpatient and outpatient care, routine haematology and biochemistry, and microscopy for parasitological confirmation of rHAT. Rumphi District Hospital serves a rural catchment bordering Vwaza Marsh Wildlife Reserve, a recognised focus of T. brucei rHAT in northern Malawi; WHO Global Health Observatory data provide the national reporting framework for annual rHAT case counts.^1,4 An evaluation conducted from August 2014 to July 2017 identified 78 cases in the focus, including six deaths, and showed that bringing diagnostic services closer to at-risk communities promoted earlier diagnosis and improved outcomes.¹⁶ A subsequent One Health analysis estimated an incidence of 5.9 hospital-diagnosed human trypanosomiasis cases per 100 000 people in Rumphi District.¹⁷ These data support the continued relevance of RDH as a sentinel treatment and surveillance site.

We included all patients aged ≥ 6 years admitted to RDH (northern Malawi) with parasitologically confirmed T. brucei rHAT between 01 January 2012 and 31 December 2024, using national and WHO diagnostic criteria: trypanosomes detected by microscopy in blood or CSF. All the included patients received the same treatment, Fexinidazole. We excluded records with incomplete data defined as missing CSF WCC and predetermined covariate results at admission, Day 5 and Day 11 time-points.

This was a retrospective cohort study of all eligible admissions from 01 January 2012 to 31 December 2024. Therefore, no a priori sample-size calculation was performed. Furthermore, in line with current recommendations, we do not present post hoc power,¹⁸ because post hoc power estimates are mathematically determined by the observed effect size and p-value. They add no information beyond the reported confidence intervals (CIs), for completeness, we therefore report precision via 95% CI rather than post hoc power.¹⁸

Data were obtained from patient files, ward registers and laboratory logs using a standardised Excel spreadsheet. Clinical and laboratory data were recorded at up to two analytic time points: Baseline (Day 0) and Day 11 (end of hospitalisation). Day 5 values were variably documented and were not used for the primary outcome.

We defined all variables prior to data capturing. Age (years) was calculated at admission from the date of birth or recorded stated age; the admitting clinician recorded sex (female/male). Symptoms and examination findings were transcribed as documented (present/absent) at admission. For synthesis, we defined neurological symptoms and signs as the presence of ‘drowsiness, reduced level of awareness (lethargy), or difficulty walking; and lymphatic involvement as the presence of cervical adenomegaly and/or splenomegaly on examination’. Vital signs reflected the first set obtained at admission: heart rate (beats per minute) by pulse/monitor, systolic/diastolic blood pressure (mmHg) by automated or manual sphygmomanometer, respiratory rate (breaths per minute) by direct count, and temperature (°C) by digital thermometer per ward protocol. Functional status used the clinician-assigned Karnofsky Performance Status (0–100, higher scores indicate better function) and Glasgow Coma Scale.^{3,4,5,6,7,8,9,10,11,12,13,14,15} Routine laboratory tests were measured on same-day venous blood using the hospital’s automated analysers according to laboratory standard operating procedures: haemoglobin (g/dL), total leukocyte count (×10⁹/L), platelet count (×10⁹/L), serum creatinine (mg/dL) and serum albumin (g/dL). The primary outcome, CSF WCC, was reported as cells/μL from laboratory registers following the laboratory’s microscopy-based cell counting procedure; repeat CSF counts at Day 11 were extracted from the records. When multiple measurements were available on the same day, the value closest to the admission time was used; time-varying covariates such as haemoglobin and albumin used the concurrent value at each modelled time point. Binary variables were coded with explicit reference categories for analysis as specified in the statistical analysis section. Disease staging of rHAT followed trypanosomiasis case management guidelines (Stage 2: trypanosomes in CSF and/or CSF WCC > 5 cells/μL).¹⁹ Data were entered into a structured spreadsheet with a prespecified codebook, harmonised units and labelled categorical values (e.g., sex: female/male; Stage: 1/2). Source documents were linked by a unique study ID; direct identifiers were removed before analysis. Range and logic checks flagged implausible values for verification against source documents. Data sets and documentation were stored on a secure, password access-restricted drive.

Baseline characteristics were summarised overall. Distributional form was assessed using the Shapiro–Wilk test and graphical inspection. Normally distributed continuous variables are reported as mean (standard deviation), non-normally distributed continuous variables as median (interquartile range [IQR]), and categorical variables as number (percentage). All tests were two-sided with α = 0.05.

This study evaluated associations between routinely measured clinical and laboratory variables and cerebrospinal-fluid white-cell count; it was not designed as a clinical prediction or diagnostic-accuracy study. Discrimination, calibration, receiver operating characteristic curves, cut-off values, positive predictive values and negative predictive values were therefore not estimated. Because Stage 2 disease is partly defined by cerebrospinal-fluid white-cell count, using stage as the reference standard for receiver operating characteristic analysis would introduce incorporation bias. The findings should consequently be interpreted as hypothesis-generating correlates that require prospective validation against an independent clinical reference standard.

The primary outcome was CSF WCC (cells/μL), analysed as a count. We fitted prespecified mixed-effects negative binomial regression models with a log link and a patient-level random intercept to account for within-patient correlation across baseline and Day 11; if the negative binomial failed to converge or showed negligible over-dispersion, we used a Poisson model with robust (sandwich) standard errors as a sensitivity check. Results are reported as incidence rate ratios (IRRs) with 95% CIs and two-sided p-values. We selected candidate variables by prespecifying candidate correlates on clinical and programmatic grounds: age, sex, assessment time point (Day 11 vs. Baseline), malaria test (positive vs. negative), disease Stage (Stage 2 vs. Stage 1), haemoglobin, creatinine, Karnofsky Performance Status (per point), albumin, leukocytes, respiratory rate, facial oedema (yes vs. no) and platelets. Age, sex, time point and stage were retained in all models irrespective of statistical significance. Remaining variables entered the multivariable model if their univariable mixed-effects count regression met p < 0.05. Time-varying covariates used values measured at each time point. Disease Stage (2 vs. 1) was entered as a fixed covariate to establish criterion validity (Stage 2 ↔ higher CSF WCC). We did not specify stage as a primary outcome because its definition incorporates CSF WCC (incorporation bias). Instead, modelling CSF WCC as a count preserved information, enabled mixed-effects inference across baseline and Day 11, and avoided loss of power from dichotomisation. Where multiple correlation screens were performed, we present both raw and Benjamini–Hochberg false discovery rate (BH-FDR)–adjusted p-values.^20,21

We assessed over-dispersion and compared negative-binomial with Poisson specifications Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), likelihood-ratio tests), when the negative-binomial alpha approached zero, or convergence warnings arose, we fitted Poisson models with robust (sandwich) standard errors as a sensitivity check. Intraclass correlation coefficients showed substantial clustering (empty model Intraclass Correlation Coefficient [ICC] ≈ 0.45) and supported random effects after adjustment (conditional ICC ≈ 0.98; random-intercept variance 3.42). Variance-inflation factors flagged collinearity for visit day and Karnofsky (Variance-inflation Factor [VIF] > 5); sensitivity analyses excluding these terms did not change inference. Residual and influence checks did not indicate mis-specification. All tests were two-sided (α = 0.05), and we report incidence rate ratios with 95% CIs and p-values. Except for sex and respiratory rate, completeness exceeded 98%, so we used complete-case analysis; all analyses were performed in R (v4.2.3).

This retrospective study analysed anonymised hospital data and involved no direct contact with patients. Access to source hospital records was authorised by Rumphi District Hospital. The study protocol was reviewed and approved by the Stellenbosch University Health Research Ethics Committee (HREC; approval S24/11/324), which granted a waiver of individual informed consent in line with national regulations. All procedures complied with the Declaration of Helsinki and applicable local governance requirements.

A total of 43 patients aged at least 6 years with parasitologically confirmed rHAT met the eligibility criteria (Figure 1). All 43 contributed admission data and 42 contributed Day-11 cerebrospinal-fluid white-cell counts. At admission, the median cerebrospinal-fluid white-cell count was 8 cells/μL (IQR: 3–59.5; range: 0–363).

Out of the 43 patients included in the analysis, 32.6% (n = 14) were female and 67.4% (n = 29) were male. The median age at presentation was 25 years (IQR: 17.5–40.5). All patients had confirmed T. brucei rhodesiense infection with baseline clinical and laboratory data available. All participants were residents of Vwaza catchment area. Most patients were admitted through the medical ward, which is consistent with routine case management pathways for rHAT at district level.

The categorical demographics and clinical characteristics are summarised in Table 1. Fever and headache were recorded in 42 (98%) and 41 (95%) patients, respectively. The composite neurological variable – drowsiness, impaired awareness or walking difficulty – was present in 39 (91%). Malaria co-infection was documented in 10 (23%), facial oedema in 12 (28%) and lymphatic involvement in 19 (44%). Because each symptom was assessed independently, percentages across symptoms are not mutually exclusive and should not be summed.

On physical examination, 28% had facial oedema and 44% had lymphoreticular signs, which included splenomegaly and cervical adenomegaly. Laboratory assessments at baseline showed anaemia (median haemoglobin 9.6 g/dL, IQR: 7.5–11.3) and hypoalbuminaemia (median albumin 2.38 g/dL, IQR: 2.1–2.6), consistent with systemic illnesses such as human immunodeficiency virus (HIV) and/or acquired immunodeficiency syndrome (AIDS) and tuberculosis (TB). Malaria co-infection was the most frequent comorbidity (23%).

All 43 patients received fexinidazole as first-line treatment. By discharge, 42 out of 43 patients (97.7%) had completed treatment, while 1 out of 43 (2.3%) died from complications judged unrelated to fexinidazole. One patient (2.3%) developed treatment non-response and received melarsoprol as second-line therapy.

The continuous baseline characteristics are presented according to their distributions. Variables summarised as mean (standard deviation) are shown in Table 2, while variables summarised as median (IQR), including cerebrospinal-fluid white-cell count, are shown in Table 3.

In univariable mixed-effects negative binomial models, cerebrospinal-fluid white-cell count was statistically associated with age, assessment time point, platelet count, malaria test result, disease stage, haemoglobin, creatinine, Karnofsky Performance Status, albumin, peripheral leukocyte count and facial oedema (all p < 0.05). There was no evidence of association at the 5% level for sex or respiratory rate. Complete effect estimates and CIs are presented in Table 4. Because these estimates are unadjusted, interpretation is based primarily on the multivariable model.

In the adjusted mixed-effects negative binomial model, the expected cerebrospinal-fluid white-cell count at Day 11 was 75% lower than at admission (adjusted incidence rate ratio [aIRR] = 0.25, 95% CI: 0.14–0.45; p < 0.001). Stage 2 disease was associated with a higher expected count than Stage 1 disease (aIRR = 4.23, 95% CI: 1.35–13.30; p = 0.013). Each 1 g/dL increase in haemoglobin and albumin was associated with 17% and 30% lower expected counts, respectively (haemoglobin aIRR = 0.83, 95% CI: 0.78–0.89; p < 0.001; albumin aIRR = 0.70, 95% CI 0.54–0.92; p = 0.009). Karnofsky Performance Status and facial oedema were also statistically associated with the outcome; these findings require cautious interpretation because of collinearity and possible residual confounding. Age, sex, malaria test result, creatinine and peripheral leukocyte count were not independently associated at the 5% level.