We evaluated the treatment failure rate among late-stage human African trypanosomiasis (HAT) patients treated with melarsoprol in Arua, northern Uganda, between September 1995 and August 1996, and identified the risk factors for treatment failure. We conducted a retrospective cohort study in October 1998, and performed a survival analysis. A treatment failure was defined as a late-stage HAT patient fully treated with melarsoprol and classified as an HAT case at any follow-up visit within 2 years after treatment. Among 428 patients treated in the study period, 130 (30·4%) were identified as treatment failure within 2 years after discharge. The multivariate analysis showed that patients who experienced treatment failure after melarsoprol were more likely to have been admitted as a relapsing case (relative hazard, RH = 11·15 [6·34–19·61]), and to have been diagnosed with trypanosomes in the lymph nodes (RH = 3·19 [2·10–4·83]) or in the cerebrospinal fluid (CSF) (RH = 1·66 [1·09–2·53]). The risk of treatment failure also increased with the number of cells in the CSF. The treatment failure rate after melarsoprol observed in Arua is greatly above the expected figures of 3–9%. More research is needed to confirm whether it is related to the variation of melarsoprol pharmacokinetics between individuals, or if it is associated with a reduced susceptibility of the trypanosomes to melarsoprol. The study emphasizes the need for second-line drugs to treat patients that have already received one or several full course(s) of melarsoprol.