Training machine learning models on patient level data segregation is crucial in practical clinical applications

This article discusses the effect of segregation of histopathology images data into three sets; training set for training machine learning model, validation set for model selection and test set for testing model performance. We found that one must be cautious when segregating histological images data (slides) into training, validation and test sets because subtle mishandling of data can introduce data leakage and gives illusively good results on the test set. We performed this study on gene mutation prediction performance by using the deep neural network in the paper of Coudray et al. [1]. By using the provided code and the same set of data, we discovered that data segregation method of the paper suffered from a data leakage problem [2]. The paper pools all the slides from all patients and then segregates them exclusively into training, validation and test sets. In this way, none of the slides is used in more than one set. This seems to be a clean separation of the data. However, the paper did not consider that some slides were strongly correlated. For example, if the tumor of a patient is cut and stained to produce multiple slides, these slides are strongly correlated. If one slide is used for training and another one is used for testing, essentially, the deep neural network can memorize the pattern on the slide in the training set and apply this memory on the slide in the test set. Hence, by memorization, the deep neural network can predict very well on the slide in the test set. This mechanism of prediction is not useful in a practical clinical setting since no two tumors are the same in the real world. In this real setting, we demand the deep neural network to generalize across patients and tumors. Hereafter, we call this way of data segregation slide-level segregation. There is a better way to perform data segregation that is compatible for deployment of deep learning model in practical clinical settings. First, the patients are segregated exclusively into training, validation and test sets. All the slides belonging to the patients in the training set are used solely for training. Similarly, all the slides belonging to the patients in the test set are used for testing only. Segregation of data in this way forces the deep neural network to generalize across patients. We call this way of data segregation patient-level segregation.In slide-level segregation approach analysis, we obtained similar results to that presented in the paper by Coudray et al. [1]: overall performance on the test set was good. However, it was illusory due to data leakage. The model gave very good testing results on the slides that come from a patient who also has slides in the training set. On the other hand, the test result was quite bad on the slides that come from a patient who does not have any slides in the training set. Hereafter, we call the slide in the test set as seen-patient data if the corresponding patient also has some slides in the training set. Otherwise, the slide in the test set is called unseen-patient data if the corresponding patient does not have slides in the training set. Furthermore, we analyzed performance of the model on the data segregated by the patient-level segregation approach. Note that, in this approach, all patients in the test set mimics the real world clinical workflow. We observed a significant drop in the performance of the model on the test set of patient-level segregation approach compared to the performance on the test set of slide-level segregation approach. Moreover, the performance of the model on the test set of patient-level segregation approach was very similar to the performance on the unseen-patients data in the test set of slide-level segregation approach. Hence, we conclude that patient-level segregation approach is crucial and appropriate to simulate real world scenario, where each patient in the test set can be thought as a patient walking into clinic tomorrow.