A physiologically-focused approach to training multi-modality AI algorithms in medicine

一种以生理学为中心的医学多模态人工智能算法训练方法

• 批准号: 10687584
• 负责人: Geoffrey H Tison
• 金额: $ 145.35万
• 依托单位: UNIVERSITY OF CALIFORNIA, SAN FRANCISCO
• 依托单位国家: 美国
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-09-06 至 2026-08-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10687584
• 关键词: Address; Algorithms; Architecture; Artificial Intelligence; Artificial Intelligence platform; Categories; Clinical; Complex; Data; Diagnosis; Diagnostic tests; Healthcare; Heart failure; Heterogeneity; Individual; Institution; Knowledge; Learning; Manufacturer; Medical; Medical Imaging; Medicine; Methodology; Modality; Outcome; Patients; Performance; Phenotype; Physicians; Physiological; Physiology; Problem Solving; Public Health; Research; Source; Training; Work; artificial intelligence algorithm; body system; clinically relevant; cohort; data format; deep neural network; human disease; improved; innovation; large scale data; medical specialties; multimodal data; multimodality; neural network algorithm; neural network architecture; novel; prospective; real world application; success; treatment planning

PROJECT SUMMARY / ABSTRACT Artificial Intelligence (AI) algorithms have demonstrated success in numerous medical applications. In particular, they excel at analyzing raw medical data formats (such as medical imaging) which has been driven largely by a category of algorithms called deep neural networks (DNN) analyzing single diagnostic tests or modalities indi- vidually. However human disease physiology is rarely confined to one organ system or wholly captured by one diagnostic test. For many of the most clinically-relevant medical applications, the limitation of only being able to analyze one data modality places a major barrier on the types and complexity of medical problems that can be solved. Nearly all important medical decisions require consideration of multiple pieces of information simultane- ously. Current state-of-the-art DNN architectures commonly used in medicine do not readily accept multiple raw data modalities and do not perform highly-complex tasks requiring differential consideration of multi-modal data. To achieve higher-level complex medical reasoning, medical AI algorithms will require fundamental physiologi- cally-focused innovation. The overall objective of this application is to establish a novel physiologically-focused approach to train AI algorithms in medicine that, if successful, will supersede existing AI approaches by func- tioning similarly to the way physician experts triangulate information from multiple sources to arrive at conclu- sions. This research is significant because it addresses the two largest methodologic barriers confronting the real-world application of AI in medicine and that are relevant to all medical specialties: data pre-processing and DNN algorithm architecture. This proposal innovates at the intersection of AI and physiology by developing a new DNN architecture that can accept and learn from multi-modality data in a manner that accommodates a priori medical and physiologic knowledge. Through this novel DNN architecture, resulting algorithms will be able to draw complementary information from multiple inter-related data modalities, similar to how an expert physician considers multiple sources of information to derive a diagnosis or treatment plan. In addition to algorithmic inno- vation, medical AI lacks a scalable approach to perform large-scale data pre-processing, given the heterogeneity of real-world medical data across a wide range of hardware manufacturers and healthcare institutions. Solving this is critical since training data is so important to developing high-performing AI algorithms. This project will also develop an automated approach to perform data pre-processing and harmonization of medical data that is modality-agnostic. To develop and refine these innovations in real-world data, both the data pre-processing pipeline and multi-modal DNN architecture will be applied and prospectively validated to identify heart failure- related phenotypes in a large multi-modality cohort of heart failure patients. The expected outcome of this pro- posal is a general-purpose AI platform that will enable multi-modal medical AI algorithms to be developed that will achieve better performance for existing tasks while also expanding the scope of medical tasks can be ac- complished through AI.

项目总结/摘要 人工智能（AI）算法在许多医疗应用中取得了成功。特别是， 他们擅长分析原始医学数据格式（如医学成像），这主要是由 一类称为深度神经网络（DNN）的算法分析单个诊断测试或模态， vidually。然而，人类疾病生理学很少局限于一个器官系统或完全由一个器官系统捕获 诊断测试对于许多最临床相关的医疗应用，仅能够 分析一种数据模态会对医疗问题的类型和复杂性造成重大障碍， 解决了几乎所有重要的医疗决策都需要同时考虑多条信息- 是的目前在医学中常用的最先进的DNN架构不容易接受多个原始数据。 数据模态，并且不执行需要对多模态数据进行区别考虑的高度复杂的任务。 为了实现更高层次的复杂医疗推理，医疗AI算法将需要基础生理学， 专注于创新。本申请的总体目标是建立一种新的以生理学为重点的药物组合物。 在医学中训练人工智能算法的方法，如果成功，将取代现有的人工智能方法， 类似于医生专家从多个来源三角测量信息以得出结论的方式， Ssions。这项研究是重要的，因为它解决了两个最大的方法论障碍所面临的 人工智能在医学中的实际应用，并与所有医学专业相关：数据预处理和 DNN算法架构。该提案在人工智能和生理学的交叉点上进行了创新， 新的DNN架构，可以接受和学习多模态数据的方式，适应 先验医学和生理学知识。通过这种新颖的DNN架构，所产生的算法将能够 从多个相互关联的数据模式中提取补充信息，类似于专家医生如何 考虑多个信息来源以得出诊断或治疗计划。除了算法创新之外， 然而，考虑到异质性，医疗AI缺乏可扩展的方法来执行大规模数据预处理。 在广泛的硬件制造商和医疗保健机构的真实世界的医疗数据。解决 这是至关重要的，因为训练数据对于开发高性能AI算法非常重要。该项目将 还开发了一种自动化方法来执行数据预处理和医疗数据的协调， 模态不可知论为了在真实世界的数据中开发和完善这些创新，数据预处理 管道和多模式DNN架构将被应用并进行前瞻性验证，以识别心力衰竭- 在一个大型多模态心力衰竭患者队列中的相关表型。这一预期结果的支持， AI是一个通用的AI平台，它将使多模态医疗AI算法能够被开发出来， 将为现有任务实现更好的性能，同时还可以扩展医疗任务的范围， 通过AI实现。