Lung Cancer Survival Prediction via Machine Learning Regression, Classification, and Statistical Techniques.

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Predictive models for lung cancer survival often struggle with accuracy for longer timeframes.
• Existing regression models show reduced performance when predicting survival beyond 6 months.

Purpose of the Study:

• To develop an improved approach for predicting lung cancer patient survival time.
• To enhance model accuracy by combining regression and classification techniques.

Main Methods:

• Utilized de-identified lung cancer patient data from the Surveillance, Epidemiology, and End Results (SEER) database.
• Employed ANOVA for feature selection, Random Forests for classification, and Linear Regression, Gradient Boosted Machines (GBM), and Random Forests for regression.
• Assessed classification accuracy using a confusion matrix and regression accuracy via Root Mean Square Error (RMSE).

Main Results:

• Random Forests (RF) demonstrated superior performance for survival times less than or equal to 6 months (RMSE 10.52) and greater than 24 months (RMSE 20.51).
• Gradient Boosted Machines (GBM) achieved the best performance for predicting survival times between 7 and 24 months (RMSE 15.65).
• Comparative analysis indicated that regression models generally perform better for shorter survival durations than RMSE values alone suggest.

Conclusions:

• Combining regression and classification models offers a more robust method for predicting lung cancer survival.
• The developed hybrid models show promise in improving the accuracy of survival time predictions across different timeframes.
• Further analysis suggests that visual comparisons may reveal model performance nuances not fully captured by RMSE, especially for shorter survival predictions.