Leveraging AI and clinical guidelines to identify undiagnosed and under-treated CKD patients for treatment optimisation

Abstract

Abstract Despite the rising prevalence and burden of chronic kidney disease (CKD), many patients remain undiagnosed, or under-treated, which negatively impacts patients’ lives and can lead to an increased strain on healthcare systems. This is partly due to time-consuming and error prone manual review and documentation, which consumes approximately 37% of clinicians’ time (1). A possible solution is to configure guideline-based artificial intelligence (AI) programmes to identify these patients in the electronic health records for clinical review, appropriate referral and treatment optimisation. This analysis aimed to evaluate whether AI, configured on Kidney Disease Improving Global Outcome (KDIGO) recommendation, could accurately find CKD patients and those at-risk of CKD compared to using ICD codes or other metadata-based searches. The AI used structured data (such as vital signs, serum creatinine & eGFR, and UACR) and unstructured data (such as clinical notes) as inputs and leveraged techniques including Natural Language Processing (NLP) and reasoning. First, an initial sample of 200 patients was manually reviewed by clinicians to establish a gold standard dataset. The AI was then evaluated on this gold standard dataset to assess its accuracy. Following this, the AI was applied to a retrospective dataset of 29,339 ICU patient records. From this larger dataset, a subset of flagged patients was manually reviewed to further evaluate AI performance, measuring the proportion of accurately classified cases. The AI configuration to final evaluation took 10 weeks. When applied to the gold standard dataset, the AI achieves 90% sensitivity and 98% precision to find CKD and at-risk CKD patients. When applied to the retrospective dataset, the AI found 2 times more CKD patients than ICD codes with 94% precision. Moreover, the AI found 6 times more at-risk CKD patients, with 90% precision. Additionally, the AI further discovered clinically valid relationships between several risk factors, such as elevated serum creatinine and congestive heart failure, helping healthcare professionals with clinical review (Figure 1). As a next step, when the AI and its dashboard are implemented in healthcare systems, healthcare professionals can assess the identified patients, including their comorbidities such as hypertension, heart failure and type 2 diabetes. This analysis provides a leading example of applying AI to identify undiagnosed at-risk people and CKD patients who are not optimally treated, with a higher performance than ICD search approaches. This solution can help reduce the clinical and economic burden on healthcare systems and provide an opportunity to healthcare professionals for optimising CKD care.Figure 1:At-risk CKD clinical features

Ähnliche Arbeiten