AUTHOR=Anley Denekew Tenaw , Akalu Temesgen Yihunie , Dessie Anteneh Mengist , Anteneh Rahel Mulatie , Zemene Melkamu Aderajew , Bayih Wubet Alebachew , Solomon Yenealem , Gebeyehu Natnael Atnafu , Kassie Gizachew Ambaw , Mengstie Misganaw Asmamaw , Abebe Endeshaw Chekol , Seid Mohammed Abdu , Gesese Molalegn Mesele , Moges Natnael , Bantie Berihun , Feleke Sefineh Fenta , Dejenie Tadesse Asmamaw , Adella Getachew Asmare , Muche Achenef Asmamaw TITLE=Prognostication of treatment non-compliance among patients with multidrug-resistant tuberculosis in the course of their follow-up: a logistic regression–based machine learning algorithm JOURNAL=Frontiers in Digital Health VOLUME=Volume 5 - 2023 YEAR=2023 URL=https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2023.1165222 DOI=10.3389/fdgth.2023.1165222 ISSN=2673-253X ABSTRACT=Introduction: Drug compliance is the act of taking medication on schedule or taking medication as prescribed and obeying other medical instructions. It is the most crucial thing in the treatment of chronic diseases particularly for multidrug resistance tuberculosis patients. Besides, it is the main reason for drug resistance and poor treatment outcome. Hence, developing the risk prediction model by using early obtainable prognostic determinants of non-compliance is vital in averting the existing, unacceptably high level of poor treatment outcomes and further drug resistance among multidrug resistance patients. Methods: A retrospective follow-up study was conducted on a total of 517 multidrug-resistant tuberculosis patients in Northwest Ethiopia. A logistic regression-based machine learning algorithm was applied to develop a risk score for the prediction of treatment non-compliance among multidrug-resistant tuberculosis patients in the selected referral hospitals of Northwest Ethiopia. The data were entered into Epi-data version 3.1 and exported to STATA version 16 and R version 4.0.5 software for analysis. A simplified risk prediction model was developed and its performance was described. It was also validated internally using the bootstrapping method. Results: Educational status, registration group (previously treated/new), treatment support, model of care, and khat use were significant prognostic features of treatment non-compliance. The model has a discriminatory power of (AUC=0.79 with a 95% CI (0.74, 0.85)) and a calibration test of (P-value = 0.5). It was internally validated by bootstrapping method and it has a relatively corrected discrimination performance (AUC of 0.78 with a 95% CI (0.73, 0.86)) with an optimism coefficient of 0.013. Conclusion: Educational status, registration group, treatment supporter, model of care, and khat use are important features that can predict treatment non-compliance of MDR-TB patients. The risk score developed has a satisfactory level of accuracy and good calibration. Besides, it is clinically interpretable and easy to use in clinical practice, for features are easily ascertainable right at patients’ enrolment time. Hence, important to reduce poor treatment outcomes and further drug resistance. Key-words: Prediction; Machine learning; Treatment compliance; Multidrug resistant tuberculosis; Ethiopia