Fig. 2: Patient-specific Bayesian calibration and prediction.

A Presents the cell and B present the volume time courses for an individual patient. The model can successfully match the measured data at V2 with quantifiable uncertainty and make a prediction for V3. The drug curves in (C, D) show the concentration time course for the slow decaying Adriamycin (left) and fast decaying cyclophosphamide (right). In (E) parameter outputs from the approximate Bayesian computation are provided as probability distribution functions (PDF). Parameters, D (diffusivity), β_A (Adriamycin decay), and β_C (cyclophosphamide decay) sample directly from the prior (e.g., no deviation from prior based on measured data), whereas the k_r,i (reduced proliferation coefficients) and α (treatment efficacy rate) find an optimal distribution other than uniform prior (e.g., parameters move to patient-specific value based on measured data). The mathematical model fits the data well based on prior information, with uncertainty provided by likelihood of parameter fits.