Sybil – Prophecies for lung cancer risk prediction?


Thoracic Oncology and Chest Procedures Network

Lung Cancer Section

The mortality benefit associated with lung cancer screening (LCS) using low dose CT (LDCT) relies, in large part, on adherence rates to annual screening of ≥90%. However, the first 1 million “real world” patients screened in the US had very low (22%) annual adherence (Silvestri, et al. Chest. 2023;S0012-3692[23]00175-7). Refining how we estimate future lung cancer risk is an important opportunity for personalized medicine to bolster adherence to follow-up after initial LDCT.

Researchers at MIT developed Sybil, a deep learning algorithm using radiomics on LDCT for LCS to accurately predict 6-year lung cancer risk (Mikhael, et al. J Clin Oncol. 2023;JCO2201345). The model was developed, trained, and tested in a total of 14,185 National Lung Screening Trial (NLST) participants including all cancer diagnoses. Within these data, Sybil’s accuracy in predicting 1-year lung cancer risk had AUC 0.92 (95% CI, 0.88-0.95) and at 6 years, AUC 0.75 (95% CI, 0.72-0.78).

The model was validated in two large independent LCS datasets, one in the US and one in Taiwan, where an LDCT can be obtained regardless of a personal smoking history. The cancer prevalence in these datasets was 3.4% and 0.9%, respectively. Reassuringly, Sybil’s performance was similar to the NLST data and was maintained in relevant subgroups such as sex, age and smoking history. Furthermore, Sybil reduced the false positive rate in the NLST to 8% at baseline scan, compared with 14% for Lung-RADS 1.0. Sybil’s algorithm, unlike others, has been made publicly available and hopefully will spur further validation and prospective study.

Robert Smyth, MD


Next Article:

Home sleep apnea test: Peripheral arterial tonometry