Machine learning (ML)-driven population pharmacokinetic (PopPK) models exhibit heterogeneous quality due to the lack of domain-specific reporting standards. To address this, we critically appraised the methodological conduct of ML-driven PopPK studies and identified domain-specific sources of bias (DSBs). Following a pre-registered protocol (https://osf.io/juvnf), quality and risk of bias (RoB) were evaluated using the PROBAST+AI tool adapted for continuous pharmacokinetic outcomes. Concurrently, DSBs were identified from recurring reporting inadequacies in the literature and mapped onto prespecified modeling requirement domains. Our appraisal included 143 individual models from 109 eligible studies. These models exhibited low quality (94.4%) and high RoB (98.6%), and the analysis domain had the highest proportion of high concerns (quality, 79.0%; RoB, 95.7%). Furthermore, we identified 15 DSBs across three domains: data quality, methodological rigor, and model robustness. Prevalent DSBs included reporting of performance measures without the inclusion of uncertainty (87.7%), failing to account for intra-individual correlation (84.7%), and data leakage in repeated-measures settings (54.4%). Overall, existing ML-driven PopPK models are subject to high RoB due to inadequately addressed modeling practices. These findings provide a foundational basis for developing domain-specific reporting standards that ensure methodological rigor for ML-driven pharmacometrics.