Scoring clinical patient notes (PNs) written by medical students is a necessary but resource-intensive task in medical education. This paper describes the organization and key lessons from a Kaggle competition on automated scoring of such notes. 1,471 teams took part in the competition and developed an extensive, publicly available code repository of varying solutions evaluated over the first public dataset for this task. The most successful approaches from this community effort are described and utilized in the development of a PN scoring system. We discuss the choice of models and system architecture with a view to operational use and scalability, and evaluate its performance on both the public Kaggle data (10 clinical cases, 43,985 PNs) and an extended internal dataset (178 clinical cases, 6,940 PNs). The results show that the system significantly outperforms a state-of-the-art existing tool for PN scoring and that task-adaptive pretraining using masked language modeling can be an effective approach even for small training samples.

This paper presents the ACTA system, which performs automated short-answer grading in the domain of high-stakes medical exams. The system builds upon previous work on neural similarity-based grading approaches by applying these to the medical domain and utilizing contrastive learning as a means to optimize the similarity metric. ACTA is evaluated against three strong baselines and is developed in alignment with operational needs, where low-confidence responses are flagged for human review. Learning curves are explored to understand the effects of training data on performance. The results demonstrate that ACTA leads to substantially lower number of responses being flagged for human review, while maintaining high classification accuracy.

This paper presents a corpus of 43,985 clinical patient notes (PNs) written by 35,156 examinees during the high-stakes USMLE® Step 2 Clinical Skills examination. In this exam, examinees interact with standardized patients - people trained to portray simulated scenarios called clinical cases. For each encounter, an examinee writes a PN, which is then scored by physician raters using a rubric of clinical concepts, expressions of which should be present in the PN. The corpus features PNs from 10 clinical cases, as well as the clinical concepts from the case rubrics. A subset of 2,840 PNs were annotated by 10 physician experts such that all 143 concepts from the case rubrics (e.g., shortness of breath) were mapped to 34,660 PN phrases (e.g., dyspnea, difficulty breathing). The corpus is available via a data sharing agreement with NBME and can be requested at https://www.nbme.org/services/data-sharing.

One of the most resource-intensive problems in the educational testing industry relates to ensuring that newly-developed exam questions can adequately distinguish between students of high and low ability. The current practice for obtaining this information is the costly procedure of pretesting: new items are administered to test-takers and then the items that are too easy or too difficult are discarded. This paper presents the first study towards automatic prediction of an item’s probability to “survive” pretesting (item survival), focusing on human-produced MCQs for a medical exam. Survival is modeled through a number of linguistic features and embedding types, as well as features inspired by information retrieval. The approach shows promising first results for this challenging new application and for modeling the difficulty of expert-knowledge questions.

Predicting the construct-relevant difficulty of Multiple-Choice Questions (MCQs) has the potential to reduce cost while maintaining the quality of high-stakes exams. In this paper, we propose a method for estimating the difficulty of MCQs from a high-stakes medical exam, where all questions were deliberately written to a common reading level. To accomplish this, we extract a large number of linguistic features and embedding types, as well as features quantifying the difficulty of the items for an automatic question-answering system. The results show that the proposed approach outperforms various baselines with a statistically significant difference. Best results were achieved when using the full feature set, where embeddings had the highest predictive power, followed by linguistic features. An ablation study of the various types of linguistic features suggested that information from all levels of linguistic processing contributes to predicting item difficulty, with features related to semantic ambiguity and the psycholinguistic properties of words having a slightly higher importance. Owing to its generic nature, the presented approach has the potential to generalize over other exams containing MCQs.