A5085004466- Artificial Intelligence in Healthcare and Education
- Machine Learning in Healthcare
- Advanced Graph Neural Networks
- Topic Modeling
- Biomedical Text Mining and Ontologies
- Radiomics and Machine Learning in Medical Imaging
- Brain Tumor Detection and Classification
- Brain Metastases and Treatment
- Complex Network Analysis Techniques
- Medical Coding and Health Information
- Stochastic Gradient Optimization Techniques
- Privacy-Preserving Technologies in Data
- Medical Imaging and Analysis
- EEG and Brain-Computer Interfaces
- Speech and dialogue systems
- Functional Brain Connectivity Studies
- Age of Information Optimization
- Medical Imaging Techniques and Applications
- Neural dynamics and brain function
- Semantic Web and Ontologies
- Intracerebral and Subarachnoid Hemorrhage Research
- Machine Learning and ELM
- Frailty in Older Adults
- Chronic Disease Management Strategies
- Natural Language Processing Techniques
NYU Langone Health
2022-2025
New York University
2022-2024
Hinge Health
2023
Carnegie Mellon University
2019-2020
Abstract Physicians make critical time-constrained decisions every day. Clinical predictive models can help physicians and administrators by forecasting clinical operational events. Existing structured data-based have limited use in everyday practice owing to complexity data processing, as well model development deployment 1–3 . Here we show that unstructured notes from the electronic health record enable training of language models, which be used all-purpose engines with low-resistance...
The adoption of large language models (LLMs) in healthcare demands a careful analysis their potential to spread false medical knowledge. Because LLMs ingest massive volumes data from the open Internet during training, they are potentially exposed unverified knowledge that may include deliberately planted misinformation. Here, we perform threat assessment simulates data-poisoning attack against Pile, popular dataset used for LLM development. We find replacement just 0.001% training tokens...
The detection and tracking of metastatic cancer over the lifetime a patient remains major challenge in clinical trials real-world care. Advances deep learning combined with massive datasets may enable development tools that can address this challenge. We present NYUMets-Brain, world's largest, longitudinal, dataset consisting imaging, follow-up, medical management 1,429 patients. Using we developed Segmentation-Through-Time, neural network which explicitly utilizes longitudinal structure...
BACKGROUND: The development of accurate machine learning algorithms requires sufficient quantities diverse data. This poses a challenge in health care because the sensitive and siloed nature biomedical information. Decentralized through federated (FL) avoid data aggregation by instead distributing to before centrally updating one global model. OBJECTIVE: To establish multicenter collaboration assess feasibility using FL train models for intracranial hemorrhage (ICH) detection without sharing...
Graph convolutional neural networks (GCNNs) are a powerful extension of deep learning techniques to graph-structured data problems. We empirically evaluate several pooling methods for GCNNs, and combinations those graph with three different architectures: GCN, TAGCN, GraphSAGE. confirm that pooling, especially DiffPool, improves classification accuracy on popular datasets find that, average, TAGCN achieves comparable or better than GCN GraphSAGE, particularly larger sparser structures.
<title>Abstract</title> The detection and tracking of metastatic cancer over the lifetime a patient remains major challenge in clinical trials real-world care. <sup>1–3</sup> Recent advances deep learning combined with massive, datasets may enable development tools that can address this challenge. We present our work NYUMets Project to develop NYUMets-Brain novel longitudinal neural network (DNN), segmentation-through-time (STT). is world's largest, longitudinal, dataset consisting imaging,...
Advances in large language models (LLMs) provide new opportunities healthcare for improved patient care, clinical decision-making, and enhancement of physician administrator workflows. However, the potential these importantly depends on their ability to generalize effectively across environments populations, a challenge often underestimated early development. To better understand reasons challenges inform mitigation approaches, we evaluated ClinicLLM, an LLM trained [HOSPITAL]'s notes,...
We present a non-invasive deep learning approach for tracking cortical spreading depressions (CSDs) in scalp electroencephalography (EEG) signals. Our method, which we refer to as CSD spatially aware convolutional network or CSD-SpArC, combines neural network, extracts temporal features from the EEG signal of each electrode, with graph exploits spatial structure signals on scalp. Using high-density EEG, this combination networks misses no CSDs, even narrowest ones (informed by widths...
Hongyi Zheng, Yixin Zhu, Lavender Jiang, Kyunghyun Cho, Eric Oermann. Proceedings of the 61st Annual Meeting Association for Computational Linguistics (Volume 4: Student Research Workshop). 2023.
Traditional evaluation metrics for classification in natural language processing such as accuracy and area under the curve fail to differentiate between models with different predictive behaviors despite their similar performance metrics. We introduce sensitivity score, a metric that scrutinizes models' at vocabulary level provide insights into disparities decision-making logic. assess score on set of representative words test using two classifiers trained hospital readmission statistics....
Recent advances in large language models have led to renewed interest natural processing healthcare using the free text of clinical notes. One distinguishing characteristic notes is their long time span over multiple documents. The unique structure creates a new design choice: when context length for model predictor limited, which part should we choose as input? Existing studies either inputs with domain knowledge or simply truncate them. We propose framework analyze sections high predictive...
INTRODUCTION: Existing clinical prediction algorithms mostly leverage small cohorts of structured data (e.g., medical imaging or laboratory data). However, large language models have demonstrated the ability to utilize unstructured outperform other machine learning approaches given sufficient data. Training on notes offers a possible alternative algorithm development in tasks. METHODS: An unlabeled dataset over seven million radiology reports and patient histories) was collected from four...
INTRODUCTION: Surgical research demands the development of clinical registries, often through time-intensive manual chart review. Natural language processing (NLP) may accelerate registry development, and an ideal automatic (autoregistry) algorithm would be highly accurate while requiring minimal data annotation. NLP approaches including bespoke Regular Expression (RegEx) classifiers Large Language Models (LLM) possess distinct strengths weaknesses have not been compared in setting...
Packing and shuffling tokens is a common practice in training auto-regressive language models (LMs) to prevent overfitting improve efficiency. Typically documents are concatenated chunks of maximum sequence length (MSL) then shuffled. However setting the atom size, for each data chunk accompanied by random shuffling, MSL may lead contextual incoherence due from different being packed into same chunk. An alternative approach utilize padding, another packing strategy, avoid only including one...
Large language models (LLMs) have recently emerged as powerful tools, finding many medical applications. LLMs' ability to coalesce vast amounts of information from sources generate a response-a process similar that human expert-has led see potential in deploying LLMs for clinical use. However, medicine is setting where accurate reasoning paramount. Many researchers are questioning the effectiveness multiple choice question answering (MCQA) benchmarks, frequently used test LLMs. Researchers...
INTRODUCTION: The development of accurate and generalizable machine learning algorithms requires sufficient quantities diverse data. This poses a challenge in healthcare due to the sensitive siloed nature biomedical information. Decentralized through federated (FL) paradigm avoid need for data aggregation by instead distributing itself before centrally updating one global model. METHODS: Five academic neurosurgery departments across US collaborated establish network using computed tomography...
Zihao Yang, Chenkang Zhang, Muru Wu, Xujin Liu, Lavender Jiang, Kyunghyun Cho, Eric Oermann. Proceedings of the 61st Annual Meeting Association for Computational Linguistics (Volume 4: Student Research Workshop). 2023.
Graph neural networks (GNNs) extend convolutional (CNNs) to graph based data. A question that arises is how much performance improvement does the underlying structure in GNN provide over CNN (that ignores this structure). To address question, we introduce edge entropy and evaluate good an indicator it for possible of GNNs CNNs. Our results on node classification with synthetic real datasets show lower values predict larger expected gains CNNs, and, conversely, higher leads smaller gains.
Graph neural networks (GNNs) extend convolutional (CNNs) to graph-based data. A question that arises is how much performance improvement does the underlying graph structure in GNN provide over CNN (that ignores this structure). To address question, we introduce edge entropy and evaluate good an indicator it for possible of GNNs CNNs. Our results on node classification with synthetic real datasets show lower values predict larger expected gains CNNs, and, conversely, higher leads smaller gains.
Graph convolutional neural networks (GCNNs) are a powerful extension of deep learning techniques to graph-structured data problems. We empirically evaluate several pooling methods for GCNNs, and combinations those graph with three different architectures: GCN, TAGCN, GraphSAGE. confirm that pooling, especially DiffPool, improves classification accuracy on popular datasets find that, average, TAGCN achieves comparable or better than GCN GraphSAGE, particularly larger sparser structures.