A dynamic and machine learning-powered clinical decision support system to enhance patient management: an example from atezolizumab in non-small cell lung cancer patients
Anna Fochesato1, Francois Mercier2, Karla Diaz-Ordaz3, Candice Jamois1
1Roche Pharma Research and Early Development, Roche Innovation Center Basel, 2Genentech Research and Early Development, Roche Innovation Center Basel, 3Department of Statistical Science, University College London
Background: Although immunotherapy has demonstrated efficacy against several cancer indications, response rates exhibit significant variability among patients [1]. Model-based clinical decision support systems (CDSS) can help oncologists assess individual benefit-risk ratios, facilitating informed actions about treatment discontinuation [2]. While many of these tools have exploited baseline information only, there is interest in utilizing early on-treatment readouts of pharmacodynamic biomarkers to disentangle initial response signals and strengthen long-term predictions [3]. Objectives: Primary objective: to develop and validate a pipeline for early predictions of survival likelihood at 1 and 2 years from treatment start in patients who received atezolizumab (ATZ) (at registered dosing regimen) as single or combination agents. Secondary objective: to validate the add-on effect of frequently sampled laboratory biomarkers on top of the sum of longest diameter (SLD) as influential predictive factors of overall survival (OS) for patients with solid tumors. Materials & Methods: Data: data from ATZ arms of BIRCH (n = 595), FIR (n = 133), and POPLAR (n = 134) Phase II studies were used for model development. Phase III OAK study data (n = 553) was used for validation, while data from patients allocated to two ATZ + Carboplatin + Paclitaxel (ACP) arms (n = 377/319), one ATZ + Carboplatin + nab-Paclitaxel (ACnP) arm (n = 325), and one ATZ + Bevacizumab + Carboplatin + Paclitaxel (ABCP) arm (n = 368) served as test cases in combination settings. Methodology: at its core, the proposed CDSS relies on model-based dynamic predictions of overall survival (OS) made at successive landmark times, using full patient information up to the time of assessment. In details, 1) The pharmacodynamics of SLD, albumin, lactate dehydrogenase, and neutrophils until various landmark times - 6, 12, and 24 weeks - were described using nonlinear mixed effects (NLME) models in Monolix 2023R1. The respective model parameters and baseline characteristics were used for training landmark-specific Random Survival Forests (RSF) in R v4.3.1. Both calibration and discrimination metrics were used for predictive performance evaluation. 2) Inductive conformal prediction [4] was calibrated on a hold-on portion of the development set using a survival-oriented nonconformity measure, providing the CDSS with a measure of uncertainty for the model outputs. 3) An adaptive decision tree at the patient level was devised to mimic real-world clinical practice: as a new patient is recruited and begins therapy, the model evaluates whether individual longitudinal data are sufficiently informative for survival outcome prediction or if additional on-treatment data is required to achieve the desired confidence threshold to output any prediction. Updates of the survival probability are done at the next landmark time for the patients invited to continue the therapy. 4) A patient-matching optimization algorithm based on the Gower distance was developed to mitigate confounding effects at baseline between monotherapy and combination cohorts and draw counterfactual insights at the population level. Results: RSF models showed good external validation and generalizability to ATZ-based combination trials, providing the majority of patients (>76%) with favorable/unfavorable predicted prognosis as early as 6 weeks into therapy with enough confidence (alpha = 0.2). Numerically, achieved performances aligned with those of immunotherapy calculators from worldwide cancer centers [5, 6] - CDSS’s C-index at 1-year for the 6-week model: ~0.67/0.72/0.65/0.63 for ACP1/ACP2/ACnP/ABCP, respectively. The embedded feature selection engine identified fewer than 10 covariates as relevant for providing survival predictions, with NLME parameters having the highest influence as determined by SHAP values [7]. Analysis of aggregated individuals revealed good diagnostics against Kaplan-Meier estimators and highlighted how study-risk interpretation is influenced by the chosen survival time frame (crossing survival curve phenomenon). Conclusion: Our CDSS leverages routinely collected biomarkers to confidently produce long-term survival prognoses and provide digital assistance to the next therapeutic steps, thereby reducing patient burden and societal spending. Preliminary translation to the population level showed promising results, indicating the potential of our tool to support Go/No-Go decisions during drug development (e.g. early arm termination in basket trials).