Abstract 213P
Background
Pulmonologists have the complex task to select the optimal treatment for patients with advanced lung cancer to extend survival duration while minimizing side-effects. They do this mainly based on patient's demographic and clinical data, patient preferences and guidelines. Digitalization of healthcare makes it possible to support this treatment selection process, aiming at more personalized and precise medicine. This study introduces a clinical decision support tool, based on prediction models for survival and burden of treatment, explainable machine learning (ML) and a Graphical User Interface (GUI) for physicians.
Methods
ML models were trained on cohorts ranging in size of 89–405 lung cancer patients with stage IIIB and IV, to predict the survival probability for different treatment options, 6 weeks, 3 months, 6 months and 1 year after the start of the first treatment. Additional models were trained on the evolution of two symptom scales, dysphagia and alopecia, of the EORTC-QLQ-LC13 questionnaire to forecast the burden of treatment. Patient demographics, laboratory results, comorbidities, tumour characteristics and treatment regimens were used as features. All models combined allow the pulmonary oncologist to simulate different therapy responses via an in-house developed GUI.
Results
The classification prediction models achieved good performance results with area under the curve values ranging from 0.78 to 0.86. In practice, a physician enters a patient identifier in the GUI. Then, the tool automatically collects all required features of the patient, flagging divergent values. After selecting a treatment schedule, the model probability outcomes are depicted, as well as the importance of each feature (based on Shapley scores) which enables the pulmonologist to understand the rationale behind the ML model's predictions.
Conclusions
We have proven that models trained on real world hospital data are capable of making reliable outcome predictions. This GUI can be used in clinical practice, provided that extra data is collected and an extensive validation procedure takes place. This will enable physicians to use data-driven predictions based on patient and disease characteristics as support in their treatment decision process.
Legal entity responsible for the study
The authors.
Funding
Flanders Innovation and Entrepreneurship.
Disclosure
All authors have declared no conflicts of interest.