Anomaly Detection of Time Series Containing Tumor Volumes

Optimizing chemotherapy treatments often involves customizing drug dosages for individuals or groups with similar traits. This process relies on converting individual patient charac-teristics into mathematical parameters in a tumor model. Mathematical tumor models are thus used for treatment generation; the models are validated and tuned (i.e., their parameters are identified) based on measurements. Consequently, the accuracy of these measurements is essential, as any errors or excessive noise in the measured tumor volumes can significantly impact the model parameters. This paper uses two different approaches for filtering outliers in the time series of tumor volumes. During the first approach, we constructed an equation based on the differences in the time series, while the second approach was based on autoencoders. The calculated metrics indicated greater accuracy for the first method; nevertheless, the neural network-based approach is more universally applicable. In the future, a combination of these methods can be applied to tumor measurements before identifying the parameters of the experimental mice.