Fiber optic communication systems are the mostwidelyusedforms of telecommunication systems owing to their high speed and longdistance communication. However, failures, degradation of optical fiber, and defective components impact their reliability. This paper proposes a predictive-maintenance framework based on supervised machinelearning algorithms that predicts failures in fiber-based ISP networks. Decision Tree, Random Forest and XGBoost algorithms are implemented on Optical Line Terminal (OLT) and Optical Network Terminal (ONT) telemetry data and the ability to predict failures is evaluated. The implemented framework improves network availability, mean-time-to-repair, and service outage. Random Forest achieved the most accurate results. As a result, this framework endorses the benefits of using ML fo