The prevalence of extensively drug-resistant (XDR) and pre-extensively drug-resistant (pre-XDR) tuberculosis in China highlights the need for rapid diagnosis. Molecular methods based on the detection of resistance-conferring mutations provide promising solution for rapid diagnosis. Here, we evaluated the accuracy of using mutations in gyrA, rrs and tlyA to predict resistance to levofloxacin (LEV), amikacin (AMK) and capreomycin (CAP), among 208 clinical multidrug-resistant (MDR) Mycobacterium tuberculosis strains collected in a local hospital in Shandong province, China. A total of 131 (63.0%, 131/208) strains were detected resistance to at least one of the 3-s line drugs by drug susceptible tests (DSTs). By comparing the mutation data with the phenotypic results, we found all mutations in three genes could specifically (with specificities from 93.9% to 100%) predict resistances with sensitivities of 77.8% for gyrA (LEV), 71.4% for rrs (AMK), 53.6% for rrs (CAP), 14.3% for tlyA (CAP), 64.3% for rrs and tlyA (CAP). The combination of these mutations could predict 68.9% and 63.4% pre-XDR and XDR TB, respectively. However, the positive predictive value of rrs for CAP resistance (57.7%) and the negative predictive values of gyrA for LEV resistance (74.5%) were not satisfying. Our results supported the use of genetic mutations to predict resistance to AMK and fluoroquinolones. An algorithm that combines molecular methods and traditional DST would be valuable for detecting resistance to second-line drugs in our hospital.