The problem of timely and accurate evaluation of well integrity is becoming increasingly relevant in the context of mature field development, high wellstream water cut, and a growing number of old wells. For production casing diagnostics, geophysical methods are typically used to identify damage and determine its interval. However, high workload of field personnel hinders prompt deployment of wireline crews to survey the integrity of wells. This results in lost oil production, increased water cut, environmental risks, increased non-productive injected volumes, and reduced key economic indices. To address these challenges, a novel approach to evaluation of casing string integrity based on machine learning models has been proposed. The paper presents a procedure for application of interpretable machine learning to detect production casing leakage and provides a comparison of this approach with the ROC-AUC statistical analysis method. The novel approach integrates the LightGBM machine learning algorithm and SHAP analysis to evaluate contribution of key features to well integrity prediction and determine their threshold values. The model training was based on data from 14,318 well surveys conducted between 2000 and 2022. The results indicate that the most important features are sulfate content, solution supersaturation ratio, and water cut. The study confirms the efficiency of interpretable machine learning methods for diagnosing complex technical systems. These results show the potential for application of such models in well integrity monitoring and well workover planning. This approach can also be used in other oil and gas applications, such as prediction of various problems and optimization of well operation conditions.