Abstract 5P
Background
Despite being accurately diagnosed by clinicians, many cancer patients are not sent for genetic screening. This is largely because many patient records have incorrect or incomplete International Classification of Disease (ICD) codes, and are therefore missed for screening, therapies and clinical trials when clinicians search based on ICD codes. This study examined whether combining Artificial Intelligence (AI) with clinical knowledge to mimic a clinician's manual review process can find more patients with lung and ovarian cancer.
Methods
An AI-driven product was configured to find lung and ovarian cancer patients based on the clinical guidelines that clinicians use when manually reviewing patient records. The product was applied to two retrospective datasets: one dataset of 99 patients to find those who were diagnosed with lung cancer and second dataset of 95 patients to find those diagnosed with ovarian cancer. Clinicians manually labelled the patient records in these two datasets for lung, ovarian or no cancer recorded. Miscoded patients were those who were determined by clinician’s manual review to have lung or ovarian cancer but lacked the corresponding ICD codes. We aimed to show the advantages of AI to find such miscoded patients.
Results
The AI product achieved 85% precision and 95% sensitivity by finding 63 lung cancer patients. ICD codes achieved 96% precision and 73% sensitivity by finding 48 patients. Hence, the AI product found 31% more lung cancer patients who were miscoded. In addition, the AI product achieved 87% precision and 100% sensitivity by finding 52 ovarian cancer patients. ICD codes achieved a 97% precision and 69% sensitivity by finding 36 ovarian cancer patients. Hence, the AI product found 44% more ovarian cancer patients who were miscoded.
Conclusions
ICD code based searches are able to find patients with high precision but several patients are missed due to miscoding. This study shows that combining clinical knowledge and AI can help mimic a clinician’s manual review and find more patients who would otherwise be missed. The discovery of such patients allows them to be screened for lung and ovarian cancer and taken through the appropriate treatment involving a therapy or a clinical trial.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Pangaea Data Limited.
Funding
Pangaea Data Limited.
Disclosure
V. Gupta: Financial Interests, Personal, Full or part-time Employment: Pangaea Data Limited; Financial Interests, Personal, Ownership Interest: Pangaea Data Limited; Financial Interests, Personal, Stocks/Shares: Pangaea Data Limited; Financial Interests, Personal, Leadership Role: Pangaea Data Limited. E.N. Kayayan, P. Falakaflaki: Financial Interests, Personal, Full or part-time Employment: Pangaea Data Limited. J. Zhang: Financial Interests, Personal, Full or part-time Employment: Pangaea Data Limited; Financial Interests, Personal, Project Lead: Pangaea Data Limited; Financial Interests, Personal, Stocks/Shares: Pangaea Data Limited. G. Gupta: Financial Interests, Personal, Full or part-time Employment: Pangaea Data Limited; Financial Interests, Personal, Advisory Role: Pangaea Data Limited.