A Machine Learning Algorithm for Predicting Therapeutic Response to Anti-PD1

We read with interest the article by Trebeschi and colleagues,¹ who report an artificial intelligence (AI)–based radiomic biomarker for assessing the anti-PD1 response in patients with non-small cell lung cancer (NSCLC) and melanoma. Compared to traditional biopsy-based assays, this noninvasive contrast-enhanced computed tomography (CT) imaging radiomic could provide more detailed characterization on each cancer lesion upon immunotherapy. Given that the radiomics have been successfully used to analyze some medical imaging,² we agree that a reliable prediction radiomic model for drug-response can generally affect the precision medicine and clinical practice.

More explicit radiomics models would benefit from stronger inclusion of other sources of data. Trebeschi and colleagues¹ constructed and validated radiomic algorithms using contrast-enhanced CT images from 1055 target lesions, yet the immunotherapy response kinetics in discovery set and validation set is not satisfactory, ranging from 0.79 to 0.83 area under the curve (AUC) in NSCLC and 0.55 to 0.64 AUC in melanoma. Apart from sample size, several other factors, especially genetic heterogeneity and inflammatory microenvironment,³ should be all taken into account when interpreting these insufficient conclusions. More importantly, recent reports have revealed that the initiation of anti-PD1/PD-L1 agents could result in hyperprogressive or pseudoprogressive disease, a paradoxical kinetics accelerating tumor growth, with ∼10% incidence in patients across multiple solid tumor types.⁴ The risk of hyperprogression and pseudoprogression should be well considered and detected in clinical drug-response assessment involving cancer immunotherapy. Apart from imaging-based circulating tumor DNA (ctDNA) monitoring, other technologies relying on peripheral blood mononuclear cells monitoring or liquid biopsies should be used in combination for detection of patients with hyperprogressive disease. Given only analyzing the CT-derived imaging data, this radiomic biomarker developed by Trebeschi et al ¹ might not effectively distinguish such hyperprogressive patients, resulting in biased antitumor response prediction. In addition, as both NSCLC and melanoma are highly heterogeneous diseases, the information from multiple dimensionalities and disciplines, such as pathological examination, laboratory data, novel cutting edge discoveries of molecular profiles, and so on, are available in routine clinical settings. Thus, to create a more reliable predictive model, an integrated AI algorithm with multilayer medical sources is extremely encouraged, and to data, it is both technically and theoretically feasible. Recent development of AI, in theory, does not need handcrafted features, as the deep learning could extract high-dimensional features in unsupervised ways to achieve certain goals.⁵ Although some arguments state that combining handcrafted features with deep learning could reduce the enrolled sample size, it should be confirmed in detail by more comparative studies. Taken together, medical imaging-based radiomics provides reliable noninvasive biomarkers for predicting the immunotherapy response. These models provide illustrative examples of precision medicine and may affect treatment strategies in cancer management.