Read about the latest research that demonstrates the potential of machine learning models to accurately predict cancer survival based on initial oncologist consultations, without the need for structured data or location-specific models. This technology could help to improve cancer care by providing personalized predictions that could inform treatment decisions.
Cancer is a leading cause of death worldwide, and accurately predicting an individual patient's survival can help improve cancer care. Currently, survival rates are calculated retrospectively and categorized primarily by cancer site and histology. However, oncologists can be inaccurate when predicting an individual patient's survival prospectively, and they have trouble factoring in personal factors such as age.
Machine learning models trained on structured data have shown promise in predicting individual patient survival by using many features of a patient's particular characteristics and disease. However, the availability of structured data varies among cancer treatment centers and patients, and not all clinical data can be easily coded or categorized for extraction and analysis.
To address these limitations, researchers have turned to unstructured data in medical documents, such as text within oncologist consultation documents, and applied natural language processing (NLP) techniques. NLP models have increasingly been applied throughout medicine, including in cancer care. However, prior studies have focused on smaller, specific documents, such as radiology or pathology reports, and have not utilized neural NLP methods to predict the survival of patients with general cancer using oncologist consultation documents.
In a recent study, researchers developed and evaluated neural NLP models to predict the survival of patients with general cancer using only their initial oncologist consultation document and no other data. The researchers found that NLP models, both traditional and neural, were able to predict 6, 36, and 60-month survival with accuracy, balanced accuracy, and AUC above 0.800, and AUC above 0.900 for the best-performing models. This performance was comparable with or superior to that of prior work, which has predicted survival only for specific types of cancer or user data that are more difficult to obtain.
One of the strengths of this study is that it utilized a common document without structured data, making it more widely applicable to different cancer treatment centers and patient populations. The researchers also found that minimal text processing was sufficient for the neural models, suggesting that understanding how words relate to each other in a document may not be as important as the presence of specific words.
However, further validation and improvement are needed before the methodology can be used in clinical practice. Future work could investigate the addition of structured data, use different models or configurations, and fine-tune the models with a relatively small number of different jurisdiction documents.
Despite these limitations, the potential for machine learning to improve cancer care by providing personalized survival predictions is a promising area for research and development. Accurately predicting an individual patient's survival could help clinicians make more informed decisions about when to refer patients to palliative care resources or consider more aggressive therapies upfront. Overall, this study highlights the potential of NLP and machine learning to improve cancer care and personalized medicine.
Share your thoughts in the comment box.
Credit: JAMA Network Open
Comments