Clinical Decision Support System based on Learning to Rank

Improving diagnostic performance with Pointwise approach to Listwise approach

Abstract

OBJECTIVES: We have developed a Clinical Decision Support System based on Learning to Rank. The purpose of this system is to support the differential diagnosis of internists and general practitioners. The system outputs a "ranked list of possible diseases" upon input of symptoms, physical findings, and results of clinical and imaging examination. Clinical Decision Support System is useful in preventing diagnostic errors. The differential diagnosis process by physicians has a high affinity with Learning to Rank. METHODS: The evaluation function of this system is NDCG. The loss function of this system is Approximate NDCG and Gumbel Approximate NDCG. The software libraries of this system are TensorFlow and TensorFlow Ranking. RESULTS: In the evaluation by the evaluation function, the diagnostic performance of this system is higher than that of the conventional methods (Pointwise approach). The inference performance of Gumbel Approximate NDCG is higher than that of conventional Approximate NDCG. The prediction score of Gumbel Approximate NDCG could be used to judge the correctness of the prediction results. In the evaluation by the real cases, this system can contribute to the prevention of diagnostic errors. CONCLUSIONS: Clinical Decision Support System is useful for physicians to assist in differential diagnosis and prevent diagnostic errors.