Long-form clinical summarization of hospital admissions has real-world significance because of its potential to help both clinicians and patients. The faithfulness of summaries is critical to their safe usage in clinical settings. To better understand the limitations of abstractive systems, as well as the suitability of existing evaluation metrics, we benchmark faithfulness metrics against fine-grained human annotations for model-generated summaries of a patient's Brief Hospital Course. We create a corpus of patient hospital admissions and summaries for a cohort of HIV patients. Annotators are asked to categorize manually highlighted summary elements into one of three categories: ``Incorrect,'' ``Missing,'' and ``Not in Notes.'' We meta-evaluate a broad set of faithfulness metrics and, across metrics, explore the importance of domain adaptation (e.g. the impact of in-domain pre-training and metric fine-tuning), the use of source-summary alignments, and the effects of distilling a single metric from an ensemble of pre-existing metrics. Off-the-shelf metrics with no exposure to clinical text correlate well yet overly rely on summary extractiveness. As a practical guide to long-form clinical narrative summarization, we find that most metrics correlate best to human judgments when provided with one summary sentence at a time and a minimal set of relevant source context.

AD and DLB co-occur frequently, but post-mortem pathology reveals limited ante-mortem accuracy in the clinical diagnosis of AD comorbid with DLB (AD/DLB) against AD alone. We tested whether a single T1-weighted (T1) MRI scan may differentiate AD/DLB and AD using heterogeneously acquired (both research and clinically acquired) neuroimaging, to maximize the amount of pathology-confirmed in-vivo MRIs. T1 MRI scans are limited to those participants with neuropathology. We examine two groups, an AD with and without LBD pathology. Slice-level 2D Convolutional neural networks (CNN) are trained to differentiate the two groups. From each scan, gray matter is segmented, normalized, smoothed and thresholded and used as input. On the subject-level, CNN records a classification accuracy of 82% and f1 score of 0.79. Prediction accuracy was higher when the scan date is closer to DOD, with a 100% accuracy recorded within a year from the date-of-death. This study demonstrates how machine learning approaches can help harmonize neuroimaging data from clinical sources to better understand disease diagnosis and progression using a true gold-standard of neuropathological confirmation. The frameworks utilized here can be extended to other diseases that are frequently co-occurring and feasibly extend to single scan diagnostic clinical utility of scans.

One of the biggest challenges in artificial intelligence for healthcare lies in the acquisition of large, accurately labeled datasets essential for deep learning (DL) model training. The scarcity of such labeled data hinders the development of generalizable models, trained on sufficiently-large datasets to perform well on unseen data. Additionally, disparities in expert opinions (for example for eye diseases like glaucoma, where there is disagreement even among clinicians on what constitutes disease) can impede establishing reliable ground truth for training. To address these issues, we use multimodal inputs: optical coherence tomography (OCT) reports and eye tracking of expert ophthalmologists as they viewed these OCT reports for glaucoma diagnosis, to enable the robust training of DL models via self-supervision derived from clinician eye movements. Eye tracking data offers a wealth of information regarding the focus of attention and the expertise level of individuals examining medical reports. In this study, we created a region-based eye movement encoding and trained a transformer model via contrastive triplet loss to learn to separate OCT reports classified as healthy from OCT reports classified as glaucoma. This approach leveraging eye tracking ‘pseudo-labels’ has the potential to enhance the performance DL models for disease diagnosis with few explicit labels.

The healthcare and life sciences sector accounts for over 30% of global data generation. Yet, approximately 97% of them remains untapped due to storage across diverse formats, as highlighted by the Deloitte report. The primary challenge is cleaning and integrating the data. To tackle this, we introduced AutoTransform, a system utilizing LLM to automatically assess, clean and transform the data. AutoTransform provides a user-friendly interface, allowing users to easily comprehend, verify, and oversee the transformation process. Once cleaned, the data can be efficiently utilized in subsequent LLM-based pipelines for tasks such as data integration and analysis.

"Glaucoma is one of the top causes of blindness worldwide. Assessing its progression is critical to determine potential visual impairment and to design sound treatment plans. Standard automated perimetry tests, commonly known as visual field (VF) tests, are clinically used to evaluate the state of functional vision. To provide an accurate and automatic diagnostic tool for clinical decision making in glaucoma progression, we utilize the predictive power of artificial intelligence (AI) and propose two vision transformer (ViT)-based deep learning (DL) networks.
First, we optimize a spatiotemporal ViT to classify a subject’s rate of glaucoma progression (GP) using only 3 baseline VFs; we explore threshold mean deviation (MD) rate of change from -0.3 to -1.5 dB/year and achieve up to 89% GP detection accuracy. Second, we develop a VF-to-VF generation architecture via a diffusion model with a ViT backbone. The model predicts future VFs with Pointwise Mean Absolute Error (PMAE) as low as 2.15 dB for mild VF deficits and is the first to extend VF prediction up to 10 years into the future. Our models are trained and validated on our ‘62K+’ dataset, the largest available of VFs to-date including at-risk, minority populations, thus ensuring our models’ generalizability. We establish our computational methods and compare testing results on the publicly available UWHVF dataset. In short, our study utilizes novel AI methods for predicting future rates and patterns of glaucoma progression in order to expedite timely treatment for better patient quality of life. The code is available at https://github.com/AI4VSLab/GP-Detection-VF-Prediction.

Medical imaging-derived tumor or lesion size quantification provides clinically relevant information for diagnosis and treatment monitoring. Multi-frequency single transducer harmonic motion imaging (ST-HMI) is an acoustic radiation force (ARF) based ultrasound elastography method that interrogates tissue mechanical properties by transmitting a multi-frequency ARF excitation pulse and estimating the induced oscillatory displacements at multiple frequencies simultaneously. In this study, an automated lesion size assessment method is presented using the U2Net model. The training-validation set contains 1094 multi-frequency ST-HMI images of heterogeneous phantoms with the inclusion’s Young’s moduli ranging from 6-70 kPa and diameters ranging from 1.7-10.5 mm. Dice score, sensitivity, specificity, and area estimation error were used to evaluate the performance of the model in the test phantom data set, in vivo 4T1 breast cancer mouse tumors, and ex vivo focused ultrasound (FUS)- induced thermal lesions. The average ± standard deviation of (Dice score, sensitivity, and specificity) was (0.92±0.04, 0.92±0.05, 0.99±0.01), and (0.87±0.07, 0.96±0.04, 0.95±0.04) in the phantoms and mouse tumors, respectively. The average absolute error in the model-predicted area of FUS-induced lesion was 9.45±0.77%. These results demonstrated the potential of multi-frequency ST-HMI to determine tumor size change in treatment monitoring and lesion size in ablation therapy using deep learning.