Divya Sharma

Importance: Transplantation is one of the few areas in medicine where the definitive treatment is rationed. Subjective decision-making pose challenges towards the transplant selection process. It has been proposed that large language models (LLMs) as artificial intelligent (AI) agents could provide objectivity in decision-making to solve complex problems. Objective: To examine the performance of a multidisciplinary selection committee of AI agents (AI-SC) as a proof-of-concept towards objectivity in the liver transplant (LT) selection process. Design: The AI-SC consisted of four LLMs: transplant hepatologist, transplant surgeon, cardiologist, and social worker. Zero-shot prompting with chain-of thought was used. Decisions were made based on clinicodemographic characteristics at time of waitlisting and LT. Setting: National LT cohort. Participants: Adult patients receiving deceased donor LT from 2004-2023 were extracted from the Scientific Registry of Transplant Recipients (SRTR) and clinical vignettes were generated. Standard absolute contraindications to LT were randomly assigned to a subset of patients to expose the AI-SC to cases of patients declined for LT. Exposures: Clinicodemographic characteristics at waitlisting and transplantation. Main Outcomes and Measures: The AI-SC’s accuracy with either: 1) listing candidates if LT would offer a 6-month or 1-year survival benefit or 2) declining candidates if contraindications to LT are present or if LT would not offer those survival benefits. Results: Of 8,412 patients, 83.6% were waitlisted and 16.4% had contraindications to LT. The AI-SC was able to accurately identify contraindications to LT (accuracy: 98.2%, 95%CI 97.9%-98.4%), predict 6-month (94.9%, 95%CI 94.4%-95.3%) and 1-year (92.0%, 95%CI 91.4%-92.6%) survival. HCC burden beyond Milan criteria was the most common reason for accepted patients who were declined by AI-SC (False Negative). Malignancy was the most common cause of death prior to 6-month or 1-year end points (False Positive). The AI-SC most frequently did not perceive a lack of social support or severe cardiopulmonary disease as barriers to LT. Conclusions and Relevance: LLMs can be leveraged to simulate the LT-SC meetings and provide accurate, objective insights on patients who may or may not benefit from LT. Lessons learned from this proof-of-concept are a provocative step towards making the LT selection process more equitable and objective. Key Points: Question: Can a multidisciplinary selection committee of artificial intelligence-based agents (AI-SC) accurately select liver transplant (LT) candidates based on potential survival benefit and contraindications to LT? Findings: Clinical vignettes were generated from 8,412 LT candidates from the Scientific Registry of Transplant Recipients (SRTR). Of these, 16.4% were randomly assigned standard absolute contraindications to LT. The AI-SC (GPT-4, OpenAI) reviewed and selected LT candidates with accuracies of 98.2% in identifying contraindications to LT, 94.9% in predicting 6-month survival benefit, and 92.0% in predicting 1-year survival benefit. Meaning: Multi-agent models may be leveraged to provide guidance towards objective decision-making in transplant candidacy.

Research is improving our understanding of how the microbiome interacts with the human body and its impact on human health. Existing machine learning methods have shown great potential in discriminating healthy from diseased microbiome states. However, Machine Learning based prediction using microbiome data has challenges such as, small sample size, imbalance between cases and controls and high cost of collecting large number of samples. To address these challenges, we propose a deep learning framework phylaGAN to augment the existing datasets with generated microbiome data using a combination of conditional generative adversarial network (C-GAN) and autoencoder. Conditional generative adversarial networks train two models against each other to compute larger simulated datasets that are representative of the original dataset. Autoencoder maps the original and the generated samples onto a common subspace to make the prediction more accurate. Extensive evaluation and predictive analysis was conducted on two datasets, T2D study and Cirrhosis study showing an improvement in mean AUC using data augmentation by 11% and 5% respectively. External validation on a cohort classifying between obese and lean subjects, with a smaller sample size provided an improvement in mean AUC close to 32% when augmented through phylaGAN as compared to using the original cohort. Our findings not only indicate that the generative adversarial networks can create samples that mimic the original data across various diversity metrics, but also highlight the potential of enhancing disease prediction through machine learning models trained on synthetic data. https://github.com/divya031090/phylaGAN.

Research shows that human microbiome is highly dynamic on longitudinal timescales, changing dynamically with diet, or due to medical interventions. In this article, we propose a novel deep learning framework ‘phyLoSTM’, using a combination of Convolutional Neural Networks and Long Short Term Memory Networks (LSTM) for feature extraction and analysis of temporal dependency in longitudinal microbiome sequencing data along with host’s environmental factors for disease prediction. Additional novelty in terms of handling variable timepoints in subjects through LSTMs, as well as, weight balancing between imbalanced cases and controls is proposed. We simulated 100 datasets across multiple time points for model testing. To demonstrate the model’s effectiveness, we also implemented this novel method into two real longitudinal human microbiome studies: (i) DIABIMMUNE three country cohort with food allergy outcomes (Milk, Egg, Peanut and Overall) and (ii) DiGiulio study with preterm delivery as outcome. Extensive analysis and comparison of our approach yields encouraging performance with an AUC of 0.897 (increased by 5%) on simulated studies and AUCs of 0.762 (increased by 19%) and 0.713 (increased by 8%) on the two real longitudinal microbiome studies respectively, as compared to the next best performing method, Random Forest. The proposed methodology improves predictive accuracy on longitudinal human microbiome studies containing spatially correlated data, and evaluates the change of microbiome composition contributing to outcome prediction. https://github.com/divya031090/phyLoSTM. Supplementary data are available at Bioinformatics online.