Sessions

Nowadays, the application of artificial intelligence (AI) is increasingly coming to the forefront in all areas of life, including veterinary medicine. AI is capable of efficiently processing and analyzing large volumes of data, thereby providing significant support in routine clinical diagnostics. The objective of our study was to develop a predictive model that forecasts post-chemotherapy neutropenia in dogs with lymphoma based on objective clinical data. The early recognition of this condition is of paramount importance, as neutropenia can delay subsequent chemotherapy cycles, negatively impacting therapeutic efficacy. Early alerts provided by the model enable the attending veterinarian to implement targeted patient monitoring and initiate timely supportive therapy.

Our analysis included the medical records of dogs diagnosed with lymphoma that received chemotherapy at the Veterinary Hematology and Oncology Center (ÁHOK) between January 12, 2022, and October 31, 2025. To ensure model accuracy and objectivity, well-measurable, exact clinical and treatment parameters were utilized. Patient-specific data recorded included body weight (kg), relative weight change (%), and age (years), while treatment-related data comprised the chemotherapeutic agent, administered dose, cycle number, and concurrent corticosteroid therapy. The incidence, onset, and severity of hematological adverse events were determined according to the Veterinary Cooperative Oncology Group Common Terminology Criteria for Adverse Events (VCOG-CTCAE) guidelines on a scale from 0 to 2: 0 (no adverse event), 1 (mild/moderate adverse event requiring no intervention or hospitalization), and 2 (severe adverse event requiring medical intervention). To maintain objectivity, these severity grades were linked to specific laboratory reference values.

Evaluation of the predictive model's performance was conducted on data from 692 chemotherapy cycles across 85 dogs, averaged over 35 independent test runs. For severe (Grade 2) neutropenic events—set as the primary clinical target—the model achieved a sensitivity of 80.6%, representing reliable detection of severe complications (30.6/38 cases). When evaluating mild and severe cases combined (Grade 1 and Grade 2), the overall sensitivity was 64.3% (102/159 cases). The specificity of the model was 79.9%, correctly identifying more than three-quarters of complication-free (Grade 0) cycles, with a false alarm rate of only 20.1%. The positive predictive value (precision) was 48.9%—indicating that nearly every second warning issued corresponded to a true neutropenic event—while the negative predictive value of 88.3% confirms that a negative signal reliably predicts the absence of complications.