TDK conference 2024

The aim of this research was to develop a predictive model using artificial intelligence to predict chemotherapy-related adverse events based on general clinical data and routine blood test results taken during treatment. Retrospective data was collected from patients’ records treated at the Veterinary Hematology and Oncology Centre between 16 September 2023 and 30 January 2024, extracted from the Doki for Vets and LabSoftLims systems. In total, 2708 blood samples were investigated, of which 719 were directly related to chemotherapy administration, and ultimately 683 of these events were used for this research.

Adverse events were classified by severity according to the Veterinary Cooperative Oncology Group Common Terminology Criteria for Adverse Events (VCOG-CTCAE). In addition to statistical methods, a correlation matrix and clinically grounded assumptions were used to evaluate the data. These selected data were then utilised to train the machine learning model. Treatments involving seven types of cytotoxic agents (doxorubicin, vincristine, vinblastine, cyclophosphamide, CCNU, mitoxantrone, carboplatin), routine laboratory results obtained during these treatments, and general clinical data of the patients (body weight, breed, age, sex, neuter status, weight change) were collected together with the potential adverse events post-administration. The input data for the machine learning model consisted of clinical and laboratory data. The model predicted the presence of adverse events (whether an adverse event would occur or not). The model was trained on up to 90% of the data and performance evaluation was done on the remaining portion. Different events for the same animal were exclusively included either in the test set or the training set, thereby creating a separate event pool for the training and test datasets.

The best model achieved 99.6% accuracy on the training set and 84.5% accuracy on the unseen test set. We compared the performance of the working model with that of five veterinary oncologists. The clinicians were given 25 data sets used by the model and asked to make similar predictions regarding the likelihood of adverse events. These data sets were neither used for model training nor for testing, and the clinicians were not aware of other patient data. This contains more adverse events, which can make the accuracy lower according to what we have seen during training. The veterinarians achieved an average prediction accuracy of 59.4%, the model achieved 72% accuracy on this test.

The occurrence of adverse events often leads to treatment delays, making their prediction critically important. The heterogeneity of the data used was high, with many input dimensions, necessitating simplifications. Severe adverse events were relatively rare. Although the current research results are promising, a larger database with samples from a broader range (more clinics) and preferably examined prospectively could yield even better results.