Madhumita Sushil

Madhumita Sushil, Vanessa E. Kennedy, Divneet Mandair et al.

Both medical care and observational studies in oncology require a thorough understanding of a patient's disease progression and treatment history, often elaborately documented in clinical notes. Despite their vital role, no current oncology information representation and annotation schema fully encapsulates the diversity of information recorded within these notes. Although large language models (LLMs) have recently exhibited impressive performance on various medical natural language processing tasks, due to the current lack of comprehensively annotated oncology datasets, an extensive evaluation of LLMs in extracting and reasoning with the complex rhetoric in oncology notes remains understudied. We developed a detailed schema for annotating textual oncology information, encompassing patient characteristics, tumor characteristics, tests, treatments, and temporality. Using a corpus of 40 de-identified breast and pancreatic cancer progress notes at University of California, San Francisco, we applied this schema to assess the zero-shot abilities of three recent LLMs (GPT-4, GPT-3.5-turbo, and FLAN-UL2) to extract detailed oncological history from two narrative sections of clinical progress notes. Our team annotated 9028 entities, 9986 modifiers, and 5312 relationships. The GPT-4 model exhibited overall best performance, with an average BLEU score of 0.73, an average ROUGE score of 0.72, an exact-match F1-score of 0.51, and an average accuracy of 68% on complex tasks (expert manual evaluation on subset). Notably, it was proficient in tumor characteristic and medication extraction, and demonstrated superior performance in relational inference like adverse event detection. However, further improvements are needed before using it to reliably extract important facts from cancer progress notes needed for clinical research, complex population management, and documenting quality patient care.

Madhumita Sushil, Dana Ludwig, Atul J. Butte et al.

Several biomedical language models have already been developed for clinical language inference. However, these models typically utilize general vocabularies and are trained on relatively small clinical corpora. We sought to evaluate the impact of using a domain-specific vocabulary and a large clinical training corpus on the performance of these language models in clinical language inference. We trained a Bidirectional Encoder Decoder from Transformers (BERT) model using a diverse, deidentified corpus of 75 million deidentified clinical notes authored at the University of California, San Francisco (UCSF). We evaluated this model on several clinical language inference benchmark tasks: clinical and temporal concept recognition, relation extraction and medical language inference. We also evaluated our model on two tasks using discharge summaries from UCSF: diagnostic code assignment and therapeutic class inference. Our model performs at par with the best publicly available biomedical language models of comparable sizes on the public benchmark tasks, and is significantly better than these models in a within-system evaluation on the two tasks using UCSF data. The use of in-domain vocabulary appears to improve the encoding of longer documents. The use of large clinical corpora appears to enhance document encoding and inferential accuracy. However, further research is needed to improve abbreviation resolution, and numerical, temporal, and implicitly causal inference.

Shenghuan Sun, Travis Zack, Madhumita Sushil et al.

Most research studying social determinants of health (SDoH) has focused on physician notes or structured elements of the electronic medical record (EMR). We hypothesize that clinical notes from social workers, whose role is to ameliorate social and economic factors, might provide a richer source of data on SDoH. We sought to perform topic modeling to identify robust topics of discussion within a large cohort of social work notes. We retrieved a diverse, deidentified corpus of 0.95 million clinical social work notes from 181,644 patients at the University of California, San Francisco. We used word frequency analysis and Latent Dirichlet Allocation (LDA) topic modeling analysis to characterize this corpus and identify potential topics of discussion. Word frequency analysis identified both medical and non-medical terms associated with specific ICD10 chapters. The LDA topic modeling analysis extracted 11 topics related to social determinants of health risk factors including financial status, abuse history, social support, risk of death, and mental health. In addition, the topic modeling approach captured the variation between different types of social work notes and across patients with different types of diseases or conditions. We demonstrated that social work notes contain rich, unique, and otherwise unobtainable information on an individual's SDoH.

Felix J. Dorfner, Amin Dada, Felix Busch et al.

Large language models (LLMs) have shown potential in biomedical applications, leading to efforts to fine-tune them on domain-specific data. However, the effectiveness of this approach remains unclear. This study evaluates the performance of biomedically fine-tuned LLMs against their general-purpose counterparts on a variety of clinical tasks. We evaluated their performance on clinical case challenges from the New England Journal of Medicine (NEJM) and the Journal of the American Medical Association (JAMA) and on several clinical tasks (e.g., information extraction, document summarization, and clinical coding). Using benchmarks specifically chosen to be likely outside the fine-tuning datasets of biomedical models, we found that biomedical LLMs mostly perform inferior to their general-purpose counterparts, especially on tasks not focused on medical knowledge. While larger models showed similar performance on case tasks (e.g., OpenBioLLM-70B: 66.4% vs. Llama-3-70B-Instruct: 65% on JAMA cases), smaller biomedical models showed more pronounced underperformance (e.g., OpenBioLLM-8B: 30% vs. Llama-3-8B-Instruct: 64.3% on NEJM cases). Similar trends were observed across the CLUE (Clinical Language Understanding Evaluation) benchmark tasks, with general-purpose models often performing better on text generation, question answering, and coding tasks. Our results suggest that fine-tuning LLMs to biomedical data may not provide the expected benefits and may potentially lead to reduced performance, challenging prevailing assumptions about domain-specific adaptation of LLMs and highlighting the need for more rigorous evaluation frameworks in healthcare AI. Alternative approaches, such as retrieval-augmented generation, may be more effective in enhancing the biomedical capabilities of LLMs without compromising their general knowledge.

Shenghuan Sun, Travis Zack, Christopher Y. K. Williams et al.

We aimed to investigate the impact of social circumstances on cancer therapy selection using natural language processing to derive insights from social worker documentation. We developed and employed a Bidirectional Encoder Representations from Transformers (BERT) based approach, using a hierarchical multi-step BERT model (BERT-MS) to predict the prescription of targeted cancer therapy to patients based solely on documentation by clinical social workers. Our corpus included free-text clinical social work notes, combined with medication prescription information, for all patients treated for breast cancer. We conducted a feature importance analysis to pinpoint the specific social circumstances that impact cancer therapy selection. Using only social work notes, we consistently predicted the administration of targeted therapies, suggesting systematic differences in treatment selection exist due to non-clinical factors. The UCSF-BERT model, pretrained on clinical text at UCSF, outperformed other publicly available language models with an AUROC of 0.675 and a Macro F1 score of 0.599. The UCSF BERT-MS model, capable of leveraging multiple pieces of notes, surpassed the UCSF-BERT model in both AUROC and Macro-F1. Our feature importance analysis identified several clinically intuitive social determinants of health (SDOH) that potentially contribute to disparities in treatment. Our findings indicate that significant disparities exist among breast cancer patients receiving different types of therapies based on social determinants of health. Social work reports play a crucial role in understanding these disparities in clinical decision-making.

Madhumita Sushil, Atul J. Butte, Ewoud Schuit et al.

Objective: Text mining of clinical notes embedded in electronic medical records is increasingly used to extract patient characteristics otherwise not or only partly available, to assess their association with relevant health outcomes. As manual data labeling needed to develop text mining models is resource intensive, we investigated whether off-the-shelf text mining models developed at external institutions, together with limited within-institution labeled data, could be used to reliably extract study variables to conduct association studies. Materials and Methods: We developed multiple text mining models on different combinations of within-institution and external-institution data to extract social factors from discharge reports of intensive care patients. Subsequently, we assessed the associations between social factors and having a do-not-resuscitate/intubate code. Results: Important differences were found between associations based on manually labeled data compared to text-mined social factors in three out of five cases. Adopting external-institution text mining models using manually labeled within-institution data resulted in models with higher F1-scores, but not in meaningfully different associations. Discussion: While text mining facilitated scaling analyses to larger samples leading to discovering a larger number of associations, the estimates may be unreliable. Confirmation is needed with better text mining models, ideally on a larger manually labeled dataset. Conclusion: The currently used text mining models were not sufficiently accurate to be used reliably in an association study. Model adaptation using within-institution data did not improve the estimates. Further research is needed to set conditions for reliable use of text mining in medical research.

Madhumita Sushil, Simon Šuster, Walter Daelemans

Understanding the behavior of a trained network and finding explanations for its outputs is important for improving the network's performance and generalization ability, and for ensuring trust in automated systems. Several approaches have previously been proposed to identify and visualize the most important features by analyzing a trained network. However, the relations between different features and classes are lost in most cases. We propose a technique to induce sets of if-then-else rules that capture these relations to globally explain the predictions of a network. We first calculate the importance of the features in the trained network. We then weigh the original inputs with these feature importance scores, simplify the transformed input space, and finally fit a rule induction model to explain the model predictions. We find that the output rule-sets can explain the predictions of a neural network trained for 4-class text classification from the 20 newsgroups dataset to a macro-averaged F-score of 0.80. We make the code available at https://github.com/clips/interpret_with_rules.

Deepali Mishra, Chaklam Silpasuwanchai, Ashutosh Modi et al.

Medical Visual Question Answering (MedVQA) is a promising tool to assist radiologists by automating medical image interpretation through question answering. Despite advances in models and datasets, MedVQA's integration into clinical workflows remains limited. This study systematically reviews 68 publications (2018-2024) and surveys 50 clinicians from India and Thailand to examine MedVQA's practical utility, challenges, and gaps. Following the Arksey and O'Malley scoping review framework, we used a two-pronged approach: (1) reviewing studies to identify key concepts, advancements, and research gaps in radiology workflows, and (2) surveying clinicians to capture their perspectives on MedVQA's clinical relevance. Our review reveals that nearly 60% of QA pairs are non-diagnostic and lack clinical relevance. Most datasets and models do not support multi-view, multi-resolution imaging, EHR integration, or domain knowledge, features essential for clinical diagnosis. Furthermore, there is a clear mismatch between current evaluation metrics and clinical needs. The clinician survey confirms this disconnect: only 29.8% consider MedVQA systems highly useful. Key concerns include the absence of patient history or domain knowledge (87.2%), preference for manually curated datasets (51.1%), and the need for multi-view image support (78.7%). Additionally, 66% favor models focused on specific anatomical regions, and 89.4% prefer dialogue-based interactive systems. While MedVQA shows strong potential, challenges such as limited multimodal analysis, lack of patient context, and misaligned evaluation approaches must be addressed for effective clinical integration.

Madhumita Sushil, Travis Zack, Divneet Mandair et al.

Although supervised machine learning is popular for information extraction from clinical notes, creating large annotated datasets requires extensive domain expertise and is time-consuming. Meanwhile, large language models (LLMs) have demonstrated promising transfer learning capability. In this study, we explored whether recent LLMs can reduce the need for large-scale data annotations. We curated a manually-labeled dataset of 769 breast cancer pathology reports, labeled with 13 categories, to compare zero-shot classification capability of the GPT-4 model and the GPT-3.5 model with supervised classification performance of three model architectures: random forests classifier, long short-term memory networks with attention (LSTM-Att), and the UCSF-BERT model. Across all 13 tasks, the GPT-4 model performed either significantly better than or as well as the best supervised model, the LSTM-Att model (average macro F1 score of 0.83 vs. 0.75). On tasks with high imbalance between labels, the differences were more prominent. Frequent sources of GPT-4 errors included inferences from multiple samples and complex task design. On complex tasks where large annotated datasets cannot be easily collected, LLMs can reduce the burden of large-scale data labeling. However, if the use of LLMs is prohibitive, the use of simpler supervised models with large annotated datasets can provide comparable results. LLMs demonstrated the potential to speed up the execution of clinical NLP studies by reducing the need for curating large annotated datasets. This may result in an increase in the utilization of NLP-based variables and outcomes in observational clinical studies.

Madhumita Sushil, Simon Šuster, Walter Daelemans

Several previous studies on explanation for recurrent neural networks focus on approaches that find the most important input segments for a network as its explanations. In that case, the manner in which these input segments combine with each other to form an explanatory pattern remains unknown. To overcome this, some previous work tries to find patterns (called rules) in the data that explain neural outputs. However, their explanations are often insensitive to model parameters, which limits the scalability of text explanations. To overcome these limitations, we propose a pipeline to explain RNNs by means of decision lists (also called rules) over skipgrams. For evaluation of explanations, we create a synthetic sepsis-identification dataset, as well as apply our technique on additional clinical and sentiment analysis datasets. We find that our technique persistently achieves high explanation fidelity and qualitatively interpretable rules.

Simon Šuster, Madhumita Sushil, Walter Daelemans

Memory networks have been a popular choice among neural architectures for machine reading comprehension and question answering. While recent work revealed that memory networks can't truly perform multi-hop reasoning, we show in the present paper that vanilla memory networks are ineffective even in single-hop reading comprehension. We analyze the reasons for this on two cloze-style datasets, one from the medical domain and another including children's fiction. We find that the output classification layer with entity-specific weights, and the aggregation of passage information with relatively flat attention distributions are the most important contributors to poor results. We propose network adaptations that can serve as simple remedies. We also find that the presence of unseen answers at test time can dramatically affect the reported results, so we suggest controlling for this factor during evaluation.

Madhumita Sushil, Simon Šuster, Kim Luyckx et al.

We have three contributions in this work: 1. We explore the utility of a stacked denoising autoencoder and a paragraph vector model to learn task-independent dense patient representations directly from clinical notes. To analyze if these representations are transferable across tasks, we evaluate them in multiple supervised setups to predict patient mortality, primary diagnostic and procedural category, and gender. We compare their performance with sparse representations obtained from a bag-of-words model. We observe that the learned generalized representations significantly outperform the sparse representations when we have few positive instances to learn from, and there is an absence of strong lexical features. 2. We compare the model performance of the feature set constructed from a bag of words to that obtained from medical concepts. In the latter case, concepts represent problems, treatments, and tests. We find that concept identification does not improve the classification performance. 3. We propose novel techniques to facilitate model interpretability. To understand and interpret the representations, we explore the best encoded features within the patient representations obtained from the autoencoder model. Further, we calculate feature sensitivity across two networks to identify the most significant input features for different classification tasks when we use these pretrained representations as the supervised input. We successfully extract the most influential features for the pipeline using this technique.

Madhumita Sushil, Simon Šuster, Kim Luyckx et al.

We have two main contributions in this work: 1. We explore the usage of a stacked denoising autoencoder, and a paragraph vector model to learn task-independent dense patient representations directly from clinical notes. We evaluate these representations by using them as features in multiple supervised setups, and compare their performance with those of sparse representations. 2. To understand and interpret the representations, we explore the best encoded features within the patient representations obtained from the autoencoder model. Further, we calculate the significance of the input features of the trained classifiers when we use these pretrained representations as input.