Adverse drug reactions associated with anti-tubercular therapy: A proof-of-concept pharmacovigilance analysis using an open-source Python analytic pipeline

Objectives To characterize the pattern, causality, severity, and seriousness of adverse drug reactions (ADRs) in patients receiving anti-tubercular therapy (ATT); to examine factors associated with ADR seriousness; and to demonstrate, as a proof of concept, that an open-source Python analytic pipeline can provide a reproducible foundation for pharmacovigilance audits and future machine-learning applications. Material and Methods In this observational, retrospective pharmacovigilance study was conducted at a tertiary-care teaching hospital in northern India. Six hundred ADR events recorded in adults (>18 years) receiving first-line or multidrug-resistant (MDR)/ extensively drug-resistant (XDR) -TB regimens were analyzed. Causality was assessed using the Naranjo algorithm and the World Health Organization-Uppsala Monitoring Centre (WHO-UMC) scale; severity was graded on the Modified Hartwig–Siegel scale; seriousness was classified per International Council for Harmonisation (ICH) E2A. Inter-scale causality agreement was quantified using Cohen’s κ. Categorical associations were tested by chi-square with Monte Carlo simulation (9,999 resamples) or Fisher’s exact test, where expected cell counts were <5; non-normally distributed continuous variables were compared using the Mann-Whitney U test with rank-biserial effect size. A post-hoc 2 × 2 analysis compared linezolid with other agents. All analyses were executed in Python 3 (Pandas, SciPy, Matplotlib); α = 0.05. Results Peripheral neuropathy was the most frequent ADR (165/600; 27.50%), followed by skin discoloration (73/600; 12.17%) and gastritis (43/600; 7.17%). The great majority of reactions were non-serious (577/600; 96.17%) and mild in severity (585/600; 97.50%); 23 (3.83%) were serious, and 6 (1.00%) severe on Hartwig grading. Causality was ‘Possible’ in 93.67% (Naranjo) and 98.33% (WHO-UMC); raw agreement was 94.00%, but Cohen’s κ was 0.23 (95% CI −0.01 to 0.47), consistent with fair agreement, and McNemar’s test was significant (χ 2 = 20.25; p <0.001), indicating systematic directional disagreement. Drug type was significantly associated with seriousness (χ 2 = 34.19; simulated p <0.001); linezolid accounted for 14/23 serious ADRs with a post-hoc odds ratio of 7.8 (95% CI 3.3–18.6) versus all other agents combined, whereas clofazimine (n = 109 exposures) contributed none. Treatment duration was significantly longer in the serious ADR group (median 270 days, IQR 210–360 vs 180, IQR 120–240; U = 9,502.5; p <0.001; r ≈ 0.42, moderate effect). Age, body weight, and sex were not significantly associated with seriousness or severity. Conclusion ADRs in this ATT cohort were frequent but predominantly mild and non-serious. Linezolid emerged as the principal drug-level factor associated with serious ADRs, supporting structured hematological and neurological surveillance in linezolid-containing regimens. The modest Cohen’s κ, despite high observed agreement, underscores the value of reporting both Naranjo and WHO-UMC assessments. A reproducible open-source Python pipeline supported the full pharmacovigilance workflow and provides an extensible foundation on which supervised ML models may subsequently be built.