AI to characterize patients with lung or ovarian cancer and those with cachexia for discovery of undiagnosed, misdiagnosed, and miscoded patients at scale in a privacy-preserving manner.

Abstract

10627 Background: Discovering undiagnosed, misdiagnosed or miscoded patients is critical to ensure suitability for clinical care pathways and for maximizing clinical trial recruitment. Meta-data (e.g. ICD) and generic NLP based approaches fall short due to: lack of scalability, significant miscoding of patient records and searching a pre-empted list of features without understanding their context or relationships between them. This study addresses these limitations through the application of Pangaea’s novel AI product (PIES) to characterize patients, which has discovered 4x more undiagnosed, misdiagnosed and miscoded patients in a scalable and privacy-preserving manner. Methods: PIES was configured on 235 clinically validated lung, ovarian and cachectic cancer patient records from secondary care. First, PIES determines all concepts which are reflected in a patient’s record based on a medical knowledge base. Second, PIES determines key clinical features and actionable relationships between them which help stratify the target patient population. Third, the findings from steps 1 and 2, are validated by clinicians and all feedback is used to refine these steps and evolve the results through a continuous loop. Following these, PIES’s characterization of the target patient population is applied across clinically validated datasets for evaluation and then across multiple healthcare systems behind their firewalls to ensure scalability and privacy compliance. Results: Applied to a dataset of 100 clinically validated patients, ICD codes correctly discovers 48 lung cancer patients and misses 18 patients (i.e. who had a diagnosis of lung cancer but were miscoded). By contrast, PIES discovers 63 lung cancer patients and misses only 3 patients. Similarly, ICD codes correctly discovers 36 ovarian cancer patients and misses 16 patients in a dataset of 100 clinically validated patients. PIES correctly discovers all 52 ovarian cancer patients. Separately, in a dataset of 235 clinically validated patients, ICD codes correctly discovers 39 cachectic cancer patients and misses 122 patients. By contrast, PIES discovers 146 cachectic cancer patients and misses only 15 patients. Conclusions: Pangaea’s AI-driven product for characterizing patients discovered 48% more lung cancer patients, 62% more ovarian cancer patients and 276% more cachectic cancer patients for screening, compared to ICD codes. This includes undiagnosed and misdiagnosed patients. Pangaea’s product platform permits the evolution of knowledge for characterizing different target patient populations through its application across multiple healthcare systems, without the need to move or touch the data, thereby ensuring scalability in a privacy compliant manner.