Study on ensemble model with weight allocation based on improved dung beetle optimization algorithm for screening colorectal cancer using laboratory test indicators.

Background: Early screening for colorectal cancer (CRC) is crucial for improving patient survival rates and reducing treatment costs. Screening is the initial risk assessment. Early detection means identifying asymptomatic cancers. Diagnosis is confirming malignancy. Population (stage I-II) was referred in our study with the consistently used "early-stage CRC". However, existing detection methods have certain deficiencies in terms of accuracy, sensitivity, and universality. Therefore, this study aims to develop an advanced machine learning-based approach using routine laboratory test indicators to enhance early CRC screening performance. Methods: This study first explored the classification effects of various common types of machine learning methods on CRC using laboratory test indicators. Subsequently, different integrated models were compared, and a weighted voting strategy based on an improved sine algorithm-guided dung beetle optimizer [improved sine algorithm-guided dung beetle optimizer with weighted voting (MSADBO-WV)] was proposed. Feature selection for CRC was performed on 45 features in the dataset. Results: The proposed MSADBO-WV method not only outperformed other integrated learning methods in terms of accuracy but also significantly exceeded the accuracy of ordinary machine learning methods [such as deep forest (DF)]. When the number of features was 26, the model achieved the highest accuracy, with the four evaluation indicators being 98.42%±1.53%, 98.46%±1.51%, 98.42%±1.53%, and 98.42%±1.53%, respectively. The analytical framework proposed in this study can be well used for screening CRC. Conclusions: MSADBO-WV demonstrates promising performance characteristics for CRC screening and will be further evaluated in prospective clinical validation studies to assist in early CRC screening and prevention strategies.