Performance of Cox proportional hazard models on recovering the ground truth of confounded exposure–response relationships for large‐molecule oncology drugs

Abstract

A Cox proportional hazard (CoxPH) model is conventionally used to assess exposure–response (E–R), but its performance to uncover the ground truth when only one dose level of data is available has not been systematically evaluated. We established a simulation workflow to generate realistic E–R datasets to assess the performance of the CoxPH model in recovering the E–R ground truth in various scenarios, considering two potential reasons for the confounded E–R relationship. We found that at high doses, when the pharmacological effects are largely saturated, missing important confounders is the major reason for inferring false‐positive E–R relationships. At low doses, when a positive E–R slope is the ground truth, either missing important confounders or mis‐specifying the interactions can lead to inaccurate estimates of the E–R slope. This work constructed a simulation workflow generally applicable to clinical datasets to generate clinically relevant simulations and provide an in‐depth interpretation on the E–R relationships with confounders inferred by the conventional CoxPH model.

Applying interpretable machine learning workflow to evaluate exposure–response relationships for large‐molecule oncology drugs

“…Similar to the traditional pharmacometric approaches for E–R analysis, static drug exposure parameters (e.g., PC min ) rather than the longitudinal PK data were used in the current ML analysis, and the E–R relationship identified may represent association/correlation or causation. The ML analysis conducted here is based on the causal assumptions of E–R relationship, as illustrated by the conceptual causal diagram for the confounded E–R relationship shown in the literature for large‐molecule oncology drugs 35 . Under this causal assumption, and the key assumption that all the baseline confounders affecting both PK and efficacy are included in the ML model, the effect of drug exposure (PC min ) on efficacy identified by the ML model and SHAP analysis can be causal instead of merely association/correlation.…”

“…Under this causal assumption, and the key assumption that all the baseline confounders affecting both PK and efficacy are included in the ML model, the effect of drug exposure (PC min ) on efficacy identified by the ML model and SHAP analysis can be causal instead of merely association/correlation. However, it is often unknown whether all confounders are available in the dataset to be included and fully adjusted by the ML model, and missing key confounders may bias the E–R relationship estimation as shown in our recent simulation study 35 . A future area of study is to utilize longitudinal PK data in ML models for E–R analysis, or apply other causal‐inference ML models, to further assess the causal conclusions.…”

“…The ML analysis conducted here is based on the causal assumptions of E–R relationship, as illustrated by the conceptual causal diagram for the confounded E–R relationship shown in the literature for large‐molecule oncology drugs. 35 Under this causal assumption, and the key assumption that all the baseline confounders affecting both PK and efficacy are included in the ML model, the effect of drug exposure (PC min ) on efficacy identified by the ML model and SHAP analysis can be causal instead of merely association/correlation. However, it is often unknown whether all confounders are available in the dataset to be included and fully adjusted by the ML model, and missing key confounders may bias the E–R relationship estimation as shown in our recent simulation study.…”

“…However, it is often unknown whether all confounders are available in the dataset to be included and fully adjusted by the ML model, and missing key confounders may bias the E–R relationship estimation as shown in our recent simulation study. 35 A future area of study is to utilize longitudinal PK data in ML models for E–R analysis, or apply other causal‐inference ML models, to further assess the causal conclusions.…”

A multistate modeling and simulation framework to learn dose–response of oncology drugs: Application to bintrafusp alfa in non‐small cell lung cancer

“…For example, patients with cancer cachexia suffer from a negative energy balance with high proteolytic clearance, and have reduced treatment benefit and/or shorter survival 15,16 . These phenomena makes clearance a potential predictor of tumor response and survival, and at the same time challenges the evaluation of causal exposure‐efficacy relationship 11,14,17 . The apparent concentration‐response relationship would be a combination of the causal concentration‐response relationship and the clearance‐response relationship, making its interpretation very challenging.…”