Tumor mutation burden (TMB) has emerged as a predictive marker for responsiveness to immune checkpoint blockade in multiple tumor types. As the gold standard, TMB is quantitated from whole exome data, but in a clinical setting it is generally approximated from targeted panel sequencing data. In this study, we systematically evaluate parameters that could affect the panel-based TMB (pTMB) assessment including panel size, gene content and local mutation determinants. By analysis simulated pTMB across different independent cohorts, we found that panels that based on cancer genes usually overestimate TMB, leading to misclassification of patients to receive improper therapy. This might be caused by positive selection for mutations on cancer genes and unlikely alleviate by removal of hotspots. To overcome this issue, we develop a parsimonious model that is capable of optimising pTMB estimation, with improved performance for patient stratification to clinical management. These findings may be immediately applicable for guiding accurate TMB approximation based on targeted panel sequencing data.