Artificial intelligence (AI) in cancer diagnosis and treatment was addressed at this year’s ASCO annual meeting.
Tsz Chun Bryan Wong, MPhil, presented the development by machine learning (ML) of an AI blood signature for detection of gastric cancer in 190,000 individuals in Hong Kong. This minimally invasive screening method uses routine blood tests and has a high adherence rate. Development was enabled by an extensive hospital system database, including history of medications for persistent dyspepsia, used as an identifying risk marker. The model, specific to suspicion of gastric cancer in Hong Kong, is highly predictive and cost effective.
A pilot implementation study, identifying clinical trial candidates in real time using AI predictions of treatment change, was the subject of a poster from the Dana-Farber Cancer Institute. The model was trained to predict six-month mortality and new systemic therapy within 30 days using new and prior imaging reports for patients with solid tumor genomic sequencing. Dana-Farber is conducting a randomized study to determine if their AI-supported algorithm can increase accrual to precision oncology trials by prompting oncologists that their patients are candidates.
Another poster, from Massachusetts General Hospital Cancer Center, described the development of ML algorithms incorporating patient reported outcome measures (PROM) to predict acute care events, including ED visits and hospitalizations within 30 days of an outpatient visit, among patients with solid tumors receiving intravenous therapy who completed PROM surveys. Including PROM in predictive models may improve model performance, although more work is needed to improve modeling and generalizability, and develop interventions.
An AI-based multi-modal using radiomic biomarkers predicted PD-(L)1 immune checkpoint inhibitor response within PD-L1 high/low/negative expression categories in patients with stage IV non-small cell lung cancer (NSCLC). According to a poster from the H. Lee Moffitt Cancer Center and Research Institute, the model may be used for clinical decision-making, pending larger cohorts and prospective studies, and seems to offer a significant benefit over current standard of care (PD-L1 immunohistochemistry).
Arsela Prelaj, MD, described the I3LUNG project that integrates real world and multi-omics data from patients with NSCLC treated with immunotherapy in Europe, the U.S., and Israel. An AI decision-making tool using retrospective real world data is expected in November 2023. Randomized studies will assess the human impact of shared decision-making on patient-physician communication for patients and the effect of AI on accuracy and the decision proves for clinicians.
Alexander T. Pearson, MD, PhD, described using AI for pathology image-based detection of clinically actionable genetic alterations. Digital AI biomarkers offer the advantages of improved accuracy, decreased testing complexity, decreased testing and overall costs, and decreased time to treatment initiation. As these move from research to clinical development stage, he warned not to trust an AI model that hasn’t been externally validated.
Adam P. Dicker, MD, PhD, cautioned that AI lacks a holistic view of patients, including their physical, mental, social, and spiritual aspects. The purpose of data analytics is to innovate, improve patient care, lead to more efficient healthcare operations, increase empathy, and reduce drudgery.
Dr. Lederman has no conflicts of interest to report.
Image by DrAfter123 / GettyImages
