A model artificial intelligence (AI) system that can identify two key genetic mutations responsible for many non-small cell lung cancers (NSCLC) has made accurate treatment recommendations in better than 85 percent of patients enrolled in a clinical trial, a researcher told attendees at the AACR Virtual Special Conference: Artificial Intelligence, Diagnosis, and Imaging held Jan. 13 (Abstract PR-03).
Wei Mu, PhD, an AI researcher at Moffitt Cancer Center, told the virtual meeting that the algorithm designed to detect two key genetic mutations for NSCLC was able to readily determine which of two treatment strategies matched each patient's mutation in many cases.
Wu's research involves developing machine learning models to analyze multimodal medical images for early diagnosis of cancer and aid personalized treatment decision-making. The investigator's team also included a number of researchers in China and subjects being treated at three prestigious Chinese cancer hospitals.
Insights Through Algorithms
The investigators applied model AI "machine-learning" algorithms modeled on what are called convolutional neural networks (ResCNN) in machine learning research. These are algorithms designed to mimic neural networks in the cerebral cortex that allow us to take "shortcuts" in our thinking to make rational insights without "convoluted" nonlinear efforts-a process known as neural "skipping." In machine learning, skipping facilitates differentiation in order to assign importance to areas on an image.
Wu and her colleagues used the algorithms in an effort to evaluate whether the approach could detect and use individuals' genetic mutations for program death-ligand 1 (PD-L1) and epidermal growth factor receptor (EGFR) proteins-both key biomarkers for diagnosing NSCLC-to pick the most appropriate of two treatment strategies: tyrosine kinase inhibitors (TKIs) or immune checkpoint inhibitors (ICIs).
"TKIs and ICIs have improved long-term survival for patients with advanced non-small cell lung cancer, but choosing the right treatment is challenging because the currently approved biomarkers for treatment decision are based on biopsies: activating mutations in epidermal growth factor receptor (EGFR) and expression of programmed death-ligand 1 (PD-L1) proteins," she explained.
"Biopsies are subject to false-negatives because of sampling bias, and they can change during therapy, rendering that treatment ineffective. Therefore, there is a compelling need to identify additional biomarkers that reflect the entire tumor and can be obtained non-invasively to help guide therapy choice."
The researchers tested their hypothesis that expression status of the two proteins might be captured by analyzing 18F-FDG PET/CT images with the algorithms.
Methodology & Accuracy
The investigators enrolled 837 NSCLC patients from four international institutions: Shanghai Pulmonary Hospital, the Fourth Hospital of Hebei Medical University, the Fourth Hospital of Harbin Medical University (HMU), and the Moffitt Cancer Center.
The models for EGFR and PD-L1 status were trained and tested in 429 patients, and then validated in 187 individuals with PET/CT images and clinical data of 616 patients. These were next tested using an external test cohort at HMU-65 persons with established EGFR mutation status and another group of 85 patients with established PD-L1 expression status.
Subsequently, the generated EGFR and PD-L1 deep learning scores (EGFR-DLS and PDL1-DLS) of the ResCNN models were further associated with progression-free survival (PFS) in 67 TKI-treated patients and 149 ICI-treated patients.
The EGFR-DLS and PDL1-DLS demonstrated high accuracy in EGFR mutation status prediction with under the curve status of 0.86 and 0.83, and 0.81 and PD-L1-positive status discrimination with under the curve values of 0.89, 0.84, and 0.82 for training, validation, and external test cohorts, respectively.
The researchers also examined how well patients responded to treatment and discovered responses to TKI and ICI. For patients with high EGFR-DLS, EGFR-TKI treatment was significantly associated with longer PFS regardless of PDL1-DLS (low PDL1-DLS, p=0.006; high PDL1-DLS, p=0.067). For patients with low EGFR-DLS, ICI treatment was significantly associated with longer PFS (p=0.007), in particular, those patients with high PDL1-DLS.
Therefore, the combination of EGFR-DLS and PDL1-DLS could be used as an alternative non-invasive decision support tool for NSCLC, they concluded.
According to the EGFR Resistors Lung Cancer Group, an organization for patients with the mutation, EGFR mutations are most common in people with lung adenocarcinoma, in non-smokers, and among women. A mutation in a gene coding for EGFR is the most common genetic change for which there are treatments available that directly target lung cancer cells.
Kurt Samson is a contributing writer.