Informed Attentive Predictors: A Generalisable Architecture for Prior Knowledge-Based Assisted Diagnosis of Cancers

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Cancer prediction and prognosis are critical due to high mortality rates.
• Existing machine learning models often require extensive data and ignore complex genetic interactions.
• Many current models lack transparency and safety for clinical use.

Purpose of the Study:

• To address limitations of current machine learning cancer predictors.
• To introduce a generalizable informed machine learning architecture, the Informed Attentive Predictor (IAP).
• To enable the use of prior knowledge (PrK) in cancer prediction models.

Main Methods:

• Developed the Informed Attentive Predictor (IAP) architecture.
• Implemented IAP to integrate prior knowledge into machine learning decision-making.
• Evaluated IAP performance on six TCGA cancer datasets.

Main Results:

• IAP models demonstrated noticeable improvements in accuracy, f1-scores, and recall rates.
• Performance gains were observed compared to non-IAP (basic predictor) counterparts.
• The architecture proved effective as an assist system framework for clinical applications.

Conclusions:

• The Informed Attentive Predictor (IAP) offers a robust framework for cancer prediction.
• Integrating prior knowledge enhances the performance and safety of machine learning predictors.
• IAP shows significant potential for improving cancer diagnosis and prognosis in clinical settings.