Segmentation of patients with small cell lung cancer into responders and non-responders using the optimal...

Area of Science:

• Oncology
• Machine Learning
• Biostatistics

Background:

• Effective cancer treatment relies on timely intervention, necessitating early identification of patients unlikely to respond to current therapies.
• Machine learning models can predict treatment response, but standard probability thresholds (e.g., 0.5) may not be optimal for patient classification.
• Accurate patient stratification is essential for personalized cancer care and optimizing treatment efficacy.

Purpose of the Study:

• To develop and validate a novel data-driven method for selecting optimal cutoff values in machine learning models for predicting cancer treatment response.
• To improve the accuracy of classifying patients as responders or non-responders compared to standard thresholding methods.
• To enhance clinical decision-making by providing a more reliable tool for treatment selection in small-cell lung cancer.

Main Methods:

• Proposed a novel data-driven approach utilizing optimal cross-validation to determine superior cutoff values for machine learning classification models.
• Applied the method to three clinical trial datasets of small-cell lung cancer patients.
• Developed a scoring system using two datasets and validated its performance on a separate test dataset.

Main Results:

• The novel method significantly improved patient segmentation compared to the standard 0.5 cutoff, with statistically significant differences in long-term survival outcomes between predicted responders and non-responders (p=0.009, HR=0.668 using Cox model).
• The standard approach failed to show a significant difference in survival between groups (p=0.194, HR=0.823 using Cox model).
• The proposed method demonstrated robust performance in distinguishing patient groups with different clinical outcomes.

Conclusions:

• The novel prediction method effectively segments new patients into responders and non-responders, offering a more accurate classification than standard approaches.
• This tool can assist clinicians in making informed decisions regarding whether a patient should continue current treatment or switch to an alternative therapy.
• The findings support the clinical utility of advanced machine learning techniques for personalized cancer treatment strategies.