Sarvesh Soni

Bushi Xiao, Sarvesh Soni, Daisy Zhe Wang

As large language models are increasingly deployed for clinical text, ensuring they can reliably signal their own uncertainty becomes critical. Most existing uncertainty quantification (UQ) methods are designed for open-domain generation and cannot localize uncertainty at the token or span level in long clinical text. We propose Reverse Probing, the first UQ framework specialized for clinical summarization, which estimates token-level uncertainty directly from pre-existing labeled summaries. Rather than sampling new outputs, Reverse Probing treats the text as a probe into the model's internal state, extracting uncertainty signals from four categories of internal activations. We evaluate on two expert-annotated clinical datasets and outperform eight adapted baselines on all metrics, achieving up to 4 times higher AUPRC while reducing inference time and computational costs. Feature analysis reveals that delta energy and neighborhood context are the most consistent predictors across all models. This study offers interpretable insights into how models internally respond to unsupported clinical content.

Sarvesh Soni, Kirk Roberts

Clinical semantic parsing (SP) is an important step toward identifying the exact information need (as a machine-understandable logical form) from a natural language query aimed at retrieving information from electronic health records (EHRs). Current approaches to clinical SP are largely based on traditional machine learning and require hand-building a lexicon. The recent advancements in neural SP show a promise for building a robust and flexible semantic parser without much human effort. Thus, in this paper, we aim to systematically assess the performance of two such neural SP models for EHR question answering (QA). We found that the performance of these advanced neural models on two clinical SP datasets is promising given their ease of application and generalizability. Our error analysis surfaces the common types of errors made by these models and has the potential to inform future research into improving the performance of neural SP models for EHR QA.

Madeline Bittner, Dina Demner-Fushman, Yasmeen Shabazz et al.

Health literacy is a critical determinant of patient outcomes, yet current screening tools are not always feasible and differ considerably in the number of items, question format, and dimensions of health literacy they capture, making documentation in structured electronic health records difficult to achieve. Automated detection from unstructured clinical notes offers a promising alternative, as these notes often contain richer, more contextual health literacy information, but progress has been limited by the lack of annotated resources. We introduce HEALIX, the first publicly available annotated health literacy dataset derived from real clinical notes, curated through a combination of social worker note sampling, keyword-based filtering, and LLM-based active learning. HEALIX contains 589 notes across 9 note types, annotated with three health literacy labels: low, normal, and high. To demonstrate its utility, we benchmarked zero-shot and few-shot prompting strategies across four open source large language models (LLMs).

Jinghui Liu, Sarvesh Soni, Anthony Nguyen

Large language models (LLMs) can generate or synthesize clinical text for a wide range of applications, from improving clinical documentation to augmenting clinical text analytics. Yet evaluations typically focus on a narrow aspect -- such as similarity or utility comparisons -- even though these aspects are complementary and best viewed in parallel. In this study, we aim to conduct a systematic evaluation of LLM-generated clinical text, which includes intrinsic, extrinsic, and factuality evaluations of synthetic clinical notes rephrased from MIMIC databases at million-note scale. Our analysis demonstrates that synthetic notes preserve core clinical information and predictive utility for coarse-grained tasks despite substantial linguistic changes, but lose fine-grained details for task like ICD coding. We show this loss of detail can be substantially mitigated by rephrasing notes by chunks rather than by the whole note, but at the cost of reduced factual precision under incomplete context. Through fact-checking and error analysis, we further find that synthesis errors are dominated by misinterpretation of clinical context, alongside temporal confusion, measurement errors, and fabricated claims. Finally, we show that the synthetic notes -- despite their task-agnostic nature -- can effectively augment task-specific training for rare ICD codes.

Sarvesh Soni, Dina Demner-Fushman

Patients have distinct information needs about their hospitalization that can be addressed using clinical evidence from electronic health records (EHRs). While artificial intelligence (AI) systems show promise in meeting these needs, robust datasets are needed to evaluate the factual accuracy and relevance of AI-generated responses. To our knowledge, no existing dataset captures patient information needs in the context of their EHRs. We introduce ArchEHR-QA, an expert-annotated dataset based on real-world patient cases from intensive care unit and emergency department settings. The cases comprise questions posed by patients to public health forums, clinician-interpreted counterparts, relevant clinical note excerpts with sentence-level relevance annotations, and clinician-authored answers. To establish benchmarks for grounded EHR question answering (QA), we evaluated three open-weight large language models (LLMs)--Llama 4, Llama 3, and Mixtral--across three prompting strategies: generating (1) answers with citations to clinical note sentences, (2) answers before citations, and (3) answers from filtered citations. We assessed performance on two dimensions: Factuality (overlap between cited note sentences and ground truth) and Relevance (textual and semantic similarity between system and reference answers). The final dataset contains 134 patient cases. The answer-first prompting approach consistently performed best, with Llama 4 achieving the highest scores. Manual error analysis supported these findings and revealed common issues such as omitted key clinical evidence and contradictory or hallucinated content. Overall, ArchEHR-QA provides a strong benchmark for developing and evaluating patient-centered EHR QA systems, underscoring the need for further progress toward generating factual and relevant responses in clinical contexts.

Sarvesh Soni, Dina Demner-Fushman

Automated approaches to answer patient-posed health questions are rising, but selecting among systems requires reliable evaluation. The current gold standard for evaluating the free-text artificial intelligence (AI) responses--human expert review--is labor-intensive and slow, limiting scalability. Automated metrics are promising yet variably aligned with human judgments and often context-dependent. To address the feasibility of automating the evaluation of AI responses to hospitalization-related questions posed by patients, we conducted a large systematic study of evaluation approaches. Across 100 patient cases, we collected responses from 28 AI systems (2800 total) and assessed them along three dimensions: whether a system response (1) answers the question, (2) appropriately uses clinical note evidence, and (3) uses general medical knowledge. Using clinician-authored reference answers to anchor metrics, automated rankings closely matched expert ratings. Our findings suggest that carefully designed automated evaluation can scale comparative assessment of AI systems and support patient-clinician communication.

Sarvesh Soni, Kirk Roberts

We apply deep learning-based language models to the task of patient cohort retrieval (CR) with the aim to assess their efficacy. The task of CR requires the extraction of relevant documents from the electronic health records (EHRs) on the basis of a given query. Given the recent advancements in the field of document retrieval, we map the task of CR to a document retrieval task and apply various deep neural models implemented for the general domain tasks. In this paper, we propose a framework for retrieving patient cohorts using neural language models without the need of explicit feature engineering and domain expertise. We find that a majority of our models outperform the BM25 baseline method on various evaluation metrics.

Sarvesh Soni, Kirk Roberts

The COVID-19 pandemic has resulted in a tremendous need for access to the latest scientific information, primarily through the use of text mining and search tools. This has led to both corpora for biomedical articles related to COVID-19 (such as the CORD-19 corpus (Wang et al., 2020)) as well as search engines to query such data. While most research in search engines is performed in the academic field of information retrieval (IR), most academic search engines$\unicode{x2013}$though rigorously evaluated$\unicode{x2013}$are sparsely utilized, while major commercial web search engines (e.g., Google, Bing) dominate. This relates to COVID-19 because it can be expected that commercial search engines deployed for the pandemic will gain much higher traction than those produced in academic labs, and thus leads to questions about the empirical performance of these search tools. This paper seeks to empirically evaluate two such commercial search engines for COVID-19, produced by Google and Amazon, in comparison to the more academic prototypes evaluated in the context of the TREC-COVID track (Roberts et al., 2020). We performed several steps to reduce bias in the available manual judgments in order to ensure a fair comparison of the two systems with those submitted to TREC-COVID. We find that the top-performing system from TREC-COVID on bpref metric performed the best among the different systems evaluated in this study on all the metrics. This has implications for developing biomedical retrieval systems for future health crises as well as trust in popular health search engines.