113P - A propensity-based analysis of SBRT and VATS for early-stage lung cancer: A guide to data-driven support and decision-making in a multidisciplinary tumour board

Background

Studies comparing video-assisted thoracic surgery (VATS) and stereotactic body radiation therapy (SBRT) for early-stage non-small cell lung cancer (NSCLC) have produced conflicting results. To cope with this scenario, artificial intelligence and machine learning processes can synthesize information from clinical data and help to construct a decision-making tool. We decided to analyse cases in our prospective database that underwent either VATS or SBRT in our institution and use these data to create an equation to predict the probability of a patient undergoing surgery or radiotherapy.

Methods

We queried our database for patients with stage IA-IIA by 8^th TNM (suspected) NSCLC referred for local treatment at the University Hospital of Leuven after a multidisciplinary tumour board (MTB) between Jan 1, 2015, and Dec 31, 2019. Baseline clinical case-mix variables and follow-up of at least 3-year survival were collected. Statistical analysis was performed with a Cox proportional hazards model for overall survival (OS) analysis and a logistic regression for the treatment predicting equation.

Results

475 patients underwent either SBRT (n = 175) or VATS (n = 307). Mean follow-up time was 4.5 years. 141 patients (30%) experienced disease recurrence: 80 of VATS group (26%) and 61 of SBRT group (35%). Eventually 164 patients died (35%) of any cause. Unadjusted OS was in favour of VATS (HR 0.37; p<0.05), but no significant survival difference was observed after propensity matching (HR 0.90; p=0.63). Based on 8 clinical features, our machine learning model was able to reproduce the MTB decision for local treatment modality with an accuracy of 84%.

Conclusions

Referral to local treatment modality should be after a careful and complete clinical assessment and MTB discussion. Artificial intelligence and machine learning can help to construct a decision-making tool. A machine learning based predictive tool for local treatment in early-stage NSCLC can combine earlier MTB expertise and clinical patient data in a real-life MTB setting. Our model needs further evaluation through test/train iterations or in external datasets.

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.