Thomas Savage

Scott L. Fleming, Alejandro Lozano, William J. Haberkorn et al. · stanford

The ability of large language models (LLMs) to follow natural language instructions with human-level fluency suggests many opportunities in healthcare to reduce administrative burden and improve quality of care. However, evaluating LLMs on realistic text generation tasks for healthcare remains challenging. Existing question answering datasets for electronic health record (EHR) data fail to capture the complexity of information needs and documentation burdens experienced by clinicians. To address these challenges, we introduce MedAlign, a benchmark dataset of 983 natural language instructions for EHR data. MedAlign is curated by 15 clinicians (7 specialities), includes clinician-written reference responses for 303 instructions, and provides 276 longitudinal EHRs for grounding instruction-response pairs. We used MedAlign to evaluate 6 general domain LLMs, having clinicians rank the accuracy and quality of each LLM response. We found high error rates, ranging from 35% (GPT-4) to 68% (MPT-7B-Instruct), and an 8.3% drop in accuracy moving from 32k to 2k context lengths for GPT-4. Finally, we report correlations between clinician rankings and automated natural language generation metrics as a way to rank LLMs without human review. We make MedAlign available under a research data use agreement to enable LLM evaluations on tasks aligned with clinician needs and preferences.

Thomas Savage, Ashwin Nayak, Robert Gallo et al.

One of the major barriers to using large language models (LLMs) in medicine is the perception they use uninterpretable methods to make clinical decisions that are inherently different from the cognitive processes of clinicians. In this manuscript we develop novel diagnostic reasoning prompts to study whether LLMs can perform clinical reasoning to accurately form a diagnosis. We find that GPT4 can be prompted to mimic the common clinical reasoning processes of clinicians without sacrificing diagnostic accuracy. This is significant because an LLM that can use clinical reasoning to provide an interpretable rationale offers physicians a means to evaluate whether LLMs can be trusted for patient care. Novel prompting methods have the potential to expose the black box of LLMs, bringing them one step closer to safe and effective use in medicine.

Seyed Amir Ahmad Safavi-Naini, Shuhaib Ali, Omer Shahab et al.

Background and Aims: This study evaluates the medical reasoning performance of large language models (LLMs) and vision language models (VLMs) in gastroenterology. Methods: We used 300 gastroenterology board exam-style multiple-choice questions, 138 of which contain images to systematically assess the impact of model configurations and parameters and prompt engineering strategies utilizing GPT-3.5. Next, we assessed the performance of proprietary and open-source LLMs (versions), including GPT (3.5, 4, 4o, 4omini), Claude (3, 3.5), Gemini (1.0), Mistral, Llama (2, 3, 3.1), Mixtral, and Phi (3), across different interfaces (web and API), computing environments (cloud and local), and model precisions (with and without quantization). Finally, we assessed accuracy using a semiautomated pipeline. Results: Among the proprietary models, GPT-4o (73.7%) and Claude3.5-Sonnet (74.0%) achieved the highest accuracy, outperforming the top open-source models: Llama3.1-405b (64%), Llama3.1-70b (58.3%), and Mixtral-8x7b (54.3%). Among the quantized open-source models, the 6-bit quantized Phi3-14b (48.7%) performed best. The scores of the quantized models were comparable to those of the full-precision models Llama2-7b, Llama2--13b, and Gemma2-9b. Notably, VLM performance on image-containing questions did not improve when the images were provided and worsened when LLM-generated captions were provided. In contrast, a 10% increase in accuracy was observed when images were accompanied by human-crafted image descriptions. Conclusion: In conclusion, while LLMs exhibit robust zero-shot performance in medical reasoning, the integration of visual data remains a challenge for VLMs. Effective deployment involves carefully determining optimal model configurations, encouraging users to consider either the high performance of proprietary models or the flexible adaptability of open-source models.

Mohammadreza Ghaffarzadeh-Esfahani, Mahdi Ghaffarzadeh-Esfahani, Arian Salahi-Niri et al.

This study compared the performance of classical feature-based machine learning models (CMLs) and large language models (LLMs) in predicting COVID-19 mortality using high-dimensional tabular data from 9,134 patients across four hospitals. Seven CML models, including XGBoost and random forest (RF), were evaluated alongside eight LLMs, such as GPT-4 and Mistral-7b, which performed zero-shot classification on text-converted structured data. Additionally, Mistral- 7b was fine-tuned using the QLoRA approach. XGBoost and RF demonstrated superior performance among CMLs, achieving F1 scores of 0.87 and 0.83 for internal and external validation, respectively. GPT-4 led the LLM category with an F1 score of 0.43, while fine-tuning Mistral-7b significantly improved its recall from 1% to 79%, yielding a stable F1 score of 0.74 during external validation. Although LLMs showed moderate performance in zero-shot classification, fine-tuning substantially enhanced their effectiveness, potentially bridging the gap with CML models. However, CMLs still outperformed LLMs in handling high-dimensional tabular data tasks. This study highlights the potential of both CMLs and fine-tuned LLMs in medical predictive modeling, while emphasizing the current superiority of CMLs for structured data analysis.

Thomas Savage, Stephen Ma, Abdessalem Boukil et al.

Large Language Model (LLM) fine tuning is underutilized in the field of medicine. Two of the most common methods of fine tuning are Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO), but there is little guidance informing users when to use either technique. In this investigation, we compare the performance of SFT and DPO for five common natural language tasks in medicine: Classification with text data, Classification with numeric data, Clinical Reasoning, Summarization, and Clinical Triage. We find that SFT alone is sufficient for Classification with text data, whereas DPO improves performance for the more complex tasks of Clinical Reasoning, Summarization and Clinical Triage. Our results establish the role and importance of DPO fine tuning within medicine, and consequently call attention to current software gaps that prevent widespread deployment of this technique.

Maia Kotelanski, Robert Gallo, Ashwin Nayak et al.

Large Language Models have difficulty communicating uncertainty, which is a significant obstacle to applying LLMs to complex medical tasks. This study evaluates methods to measure LLM confidence when suggesting a diagnosis for challenging clinical vignettes. GPT4 was asked a series of challenging case questions using Chain of Thought and Self Consistency prompting. Multiple methods were investigated to assess model confidence and evaluated on their ability to predict the models observed accuracy. The methods evaluated were Intrinsic Confidence, SC Agreement Frequency and CoT Response Length. SC Agreement Frequency correlated with observed accuracy, yielding a higher Area under the Receiver Operating Characteristic Curve compared to Intrinsic Confidence and CoT Length analysis. SC agreement is the most useful proxy for model confidence, especially for medical diagnosis. Model Intrinsic Confidence and CoT Response Length exhibit a weaker ability to differentiate between correct and incorrect answers, preventing them from being reliable and interpretable markers for model confidence. We conclude GPT4 has a limited ability to assess its own diagnostic accuracy. SC Agreement Frequency is the most useful method to measure GPT4 confidence.

Thomas Savage, Evan Madill

Recognition of respiratory distress through visual inspection is a life saving clinical skill. Clinicians can detect early signs of respiratory deterioration, creating a valuable window for earlier intervention. In this study, we evaluate whether recent advances in video transformers can enable Artificial Intelligence systems to recognize the signs of respiratory distress from video. We collected videos of healthy volunteers recovering after strenuous exercise and used the natural recovery of each participants respiratory status to create a labeled dataset for respiratory distress. Splitting the video into short clips, with earlier clips corresponding to more shortness of breath, we designed a temporal ordering challenge to assess whether an AI system can detect respiratory distress. We found a ViViT encoder augmented with Lie Relative Encodings (LieRE) and Motion Guided Masking, combined with an embedding based comparison strategy, can achieve an F1 score of 0.81 on this task. Our findings suggest that modern video transformers can recognize subtle changes in respiratory mechanics.

Nariman Naderi, Seyed Amir Ahmad Safavi-Naini, Thomas Savage et al.

This study evaluated self-reported response certainty across several large language models (GPT, Claude, Llama, Phi, Mistral, Gemini, Gemma, and Qwen) using 300 gastroenterology board-style questions. The highest-performing models (GPT-o1 preview, GPT-4o, and Claude-3.5-Sonnet) achieved Brier scores of 0.15-0.2 and AUROC of 0.6. Although newer models demonstrated improved performance, all exhibited a consistent tendency towards overconfidence. Uncertainty estimation presents a significant challenge to the safe use of LLMs in healthcare. Keywords: Large Language Models; Confidence Elicitation; Artificial Intelligence; Gastroenterology; Uncertainty Quantification

Thomas Savage

Fine-tuning methods such as Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO) have demonstrated success in training large language models (LLMs) for single-turn tasks. However, these methods fall short in multi-turn applications, such as diagnostic patient interviewing, where understanding how early conversational turns influence downstream completions and outcomes is essential. In medicine, a multi-turn perspective is critical for learning diagnostic schemas and better understanding conversation dynamics. To address this gap, I introduce Savage Conversation Forests (SCF), a reinforcement learning framework that leverages a branched conversation architecture to fine-tune LLMs for multi-turn dialogue. SCF generates multiple possible conversation continuations at each turn, enabling the model to learn how different early responses affect downstream interactions and diagnostic outcomes. In experiments simulating doctor-patient conversations, SCF with branching outperforms linear conversation architectures on diagnostic accuracy. I hypothesize that SCF's improvements stem from its ability to provide richer, interdependent training signals across conversation turns. These results suggest that a branched training architecture is an important strategy for fine tuning LLMs in complex multi-turn conversational tasks.