Detection of Adverse Event Signals with Severity Grade Classification from Cancer Patient Narrative.

Area of Science:

• Oncology
• Medical Informatics
• Natural Language Processing

Background:

• Effective adverse event (AE) management is vital for optimizing anti-cancer treatment outcomes.
• Current clinical monitoring may miss crucial AE signals, necessitating methods for continuous, real-world patient surveillance.
• Early detection of AEs can facilitate timely interventions, improving patient prognosis and quality of life.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for detecting and classifying AE signals from patient-generated text.
• To assess the performance of different DL architectures (BERT, ELECTRA, T5) in identifying AEs of varying severity grades from cancer patient blogs.
• To establish a method for early AE signal detection outside of traditional clinical settings.

Main Methods:

• Utilized a dataset of Japanese cancer patient blogs as the source for AE signal detection.
• Developed and trained three distinct DL models: BERT, ELECTRA, and T5, for classifying blog posts mentioning AEs.
• Evaluated model performance using F1 scores for classifying articles with Grade ≥ 1 and Grade ≥ 2 AEs.

Main Results:

• The T5 deep learning model achieved the highest F1 scores for both classification tasks.
• Achieved F1 scores of 0.85 for Grade ≥ 1 AE classification and 0.53 for Grade ≥ 2 AE classification.
• Demonstrated the feasibility of using DL to detect AE signals from unstructured patient text.

Conclusions:

• Deep learning models, particularly T5, can effectively detect adverse event signals from cancer patient blogs.
• This approach enables earlier detection of potential AEs, facilitating prompt medical intervention.
• Implementing such models can significantly enhance patient quality of life by improving AE management in cancer care.