PrediCare: a new diagnostic tool predicting imminent disease progression in advanced NSCLC patients by machine-learning integration of three serum biomarkers

Abstract accepted for Poster presentation during the 8th European Lung Cancer Congress (ELCC), 11-14 April 2018 in Geneva, Switzerland 

Link to Poster  

Link to Abstract

 

Yuri Kogan1, Marina Kleiman1, Shmuel Shannon1, Moran Elishmereni1, Eldad Taub1, Larisa Aptekar2, Ronen Brenner3, Raanan Berger4, Hovav Nechushtan2, Zvia Agur1

1 Optimata Ltd. Israel

2 Hadassah Medical Center, Israel

3 Wolfson Medical Center, Israel

4 Sheba Medical Center, Israel

 

Background: In advanced non-small cell lung cancer (NSCLC), most patients deteriorate rapidly and die within 1 year of diagnosis. Forecasting disease progression just prior to its clinical manifestation would allow an earlier switch to the next treatment line, thus preventing major deterioration in the patient's stature and potentially improving response to therapy. However, present serum tumor biomarkers, e.g. carcinoembryonic antigen (CEA), lack the power to signal progression. We developed PrediCare, an innovative diagnostic for continuous monitoring and alerting to forthcoming progression in late-stage NSCLC.

 

Methods: PrediCare was constructed by machine-learning modeling, and designed to process patient data throughout treatment. Data of late-stage NSCLC patients under first-line standard-of-care therapies, collected in a retrospective observational trial (NCT02577627), served for algorithm training and testing. The algorithm’s predictive ability was evaluated using diverse features of 1-3 longitudinally measured serum tumor markers (CEA, CA125, CA15.3), as pre-selected by receiver-operating-characteristic analysis. Performance was evaluated by cross-validation.

 

Results: A total of 167 NSCLC patients were assessed, the median follow-up time being 101 days. The CEA/CA125/CA15.3 combination showed statistically significant prediction ability, while the use of only 1-2 markers had lower performance. Combining the 3 markers, PrediCare accurately predicted 87/165 of the progression events (52.6% sensitivity), with 15/165 false positives (91.1% specificity). Positive predictive value, negative predictive value, accuracy and Cohen’s kappa were 68%, 85%, 81% and 0.47, respectively.

 

Conclusions: PrediCare is a new individualized medicine software tool, predicting imminent disease progression in advanced NSCLC. This improves treatment planning, and potentially increases survival. Our technology uses standard tumor biomarkers, but integrates three of them in a unique way that offers superiority over their current interpretation in the clinic. Testing of PrediCare under a larger biomarker panel is underway.