Predicting cancer survival is a crucial aspect of improving patient care, as it assists clinicians in determining treatment options and providing recommendations for end of life care.
Traditionally, cancer survival rates have been calculated retrospectively using large historical datasets categorized by factors like cancer site and tissue structure. However, these calculations often overlook personal, lifestyle, and environmental factors, leading to inaccurate predictions.
To address this limitation, machine learning (ML) — a subset of artificial intelligence (AI) — has been extensively used in medical research to develop predictive models that provide more personalized survival predictions. Unlike traditional methods, these ML models incorporate specific patient and disease characteristics, outperforming the predictions made by oncologists using conventional approaches. However, the widespread adoption of these models is limited due to variations in data availability across hospitals and patients.
Along with a team of researchers, Dr. John-Jose Nunez – a psychiatrist and Clinical Research Fellow at the University of British Columbia – has developed models that predict cancer patient survival using only the initial patient consultation notes. The models were trained on data from 47,625 patients across various cancer centres in British Columbia, achieving an accuracy of over 80% in predicting 6, 36, and 60-month survival rates. These findings were published in JAMA Network Open.
Nunez’s model is based on natural language processing, or NLP, which has recently emerged as a powerful AI branch that can comprehend complex human language. Addressing the challenges of structured data and the complexity of prognosis-influencing factors, NLP has found increasing applications in medicine and has demonstrated the ability to generate more accurate survival predictions by analyzing patient encounter documents.
Additionally, the application of neural networks, a class of ML models inspired by the interconnected nature of neurons, within NLP can enhance the understanding of human language and the relationships between specific terms.
By utilizing common medical documents such as consultation notes, these models can overcome the limitations of requiring specific data or characteristics. This type of data input also allows for the generalizability of the model across different cancer types. Furthermore, the use of AI eliminates the need for human research assistants while ensuring the confidentiality of patient records.
Since the model is trained on BC data, it presents a powerful tool for survival prediction within the province, and its reliance on consultation notes, a widely available document, makes it easily applicable across the world.
This approach addresses the challenges posed by structured data and specific characteristics, which may not be universally available for all patients. Consequently, it allows for more personalized patient care. The model can extract unique clues from consultation documents for a more individualized assessment, making it highly scalable and applicable to all cancer types—an advancement over previous models.
Looking ahead, the integration of AI in medicine holds immense potential. With the ability to analyze complex human language and extract crucial information from unstructured data, AI-driven models can revolutionize various aspects of healthcare. By leveraging large datasets, machine learning algorithms can uncover hidden patterns, identify risk factors, and optimize treatment strategies, ultimately leading to better patient outcomes.