Design an ICD-10 Prediction System for Clinical Records

Design a machine learning system that takes a clinical case record and predicts the relevant ICD-10 diagnosis codes. Explain the prediction target, how you would obtain and prepare training examples, a practical modeling approach, and how you would evaluate and serve the predictions. The case-record format, available labels, deployment scale, and required accuracy are unspecified; state your assumptions instead of assigning arbitrary targets. Explain how the design changes if a case can have several valid codes. ### What a Strong Answer Covers - A distinction between predicting recorded diagnosis codes and making a new clinical diagnosis, with an explicit single-label or multilabel formulation. - A data and label pipeline that avoids leaking the target codes into the model input and keeps related patient records out of opposing evaluation splits. - A baseline and a justified path to a richer model, including handling long records, infrequent codes, and the applicable code-set version. - Evaluation of incorrect and missing codes, confidence-based review, and traceability from a prediction to its input and model version. ```hint Define the label first A record can mention historical conditions, ruled-out conditions, and current findings. Decide what the training annotation actually marks as a target before choosing the model output. ``` ### Follow-up Questions - How would you evaluate performance on rare codes when common codes dominate the dataset? - What would you change if records regularly exceed the model's context limit?