SCH: Screening and confirmatory machine learning for explainable modeling of non-cancer deaths in cancer patients

SCH：筛选和验证机器学习，用于癌症患者非癌症死亡的可解释模型

• 批准号: 10596376
• 负责人: Lanjing Zhang
• 金额: $ 28.83万
• 依托单位: RUTGERS BIOMEDICAL AND HEALTH SCIENCES
• 依托单位国家: 美国
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-09-01 至 2026-08-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10596376
• 关键词: Address; Awareness; Cancer Model; Cancer Patient; Cancer Survivor; Cause of Death; Cessation of life; Collaborations; Colorectal Cancer; Data; Data Science; Development; Electronic Health Record; Healthcare; Human; Instruction; Knowledge; Machine Learning; Malignant Neoplasms; Malignant neoplasm of lung; Malignant neoplasm of prostate; Methods; Mission; Modeling; Morbidity - disease rate; Outcome; Patients; Physicians; Population; Population Study; Problem Solving; Public Health; Recurrence; Research; Research Personnel; Risk; Sensitivity and Specificity; Technology; Time; United States National Institutes of Health; Validation; base; cancer care; deep neural network; high school; improved; interdisciplinary collaboration; machine learning algorithm; machine learning model; malignant breast neoplasm; mortality; novel; programs; screening; tool; undergraduate student

Due to the high stakes of healthcare, the primary barrier is the extremely low tolerance of errors in healthcare practice, which requires extremely high sensitivity and specificity of any modelling. However, nearly all Machine learning (ML) models focus on improving the accuracy. It cannot yet reach both extremely high sensitivity and specificity using healthcare data. Separate screening and confirmatory ML tools are proposed to achieve very high sensitivity and specificity. Moreover, many ML algorithms suffer from the lack of clear explanations, such as deep learning and neural networks, and would unlikely meet the FAIR criteria. Cancer is the second leading cause of death in the U.S. The number of cancer survivors continues to grow; unfortunately, so does the number of non-cancer deaths in cancer patients. However, nearly all `omic and large population studies focused on binary outcomes (cancer death or recurrence). Therefore, there is an urgent need to better understand and reduce non-cancer deaths in cancer patients, using `omic and population data. To address these problems, the project will develop screening and confirmatory ML to model cancer and noncancer deaths in breast, colorectal, prostate and lung cancer patients using `omic data and electronic health records (EHR). The proposed research will result in fundamental contribution to ML tools, workflows and methods to make novel use of `omic and EHR data for cancer care. It timely meets the urgent needs in precise reduction of non-cancer deaths. This project also uniquely addresses the Transformative Data Science research theme. The interdisciplinary collaboration in this project as outlined in the Collaboration Plan will offer a diverse basis for creative problem solving and validation. The proposal has 3 broader impacts: 1) The developed novel ML algorithms and technology will enable physicians to more precisely prognosticate and treat cancer patients based on their risk of multicategory deaths. 2) The research program will support and nurture undergraduate and graduate researchers. 3) The proposed research program will support high school and undergraduate students both in the conduct of research and in awareness of ML usefulness. RELEVANCE (See instructions): The proposed research is relevant to public health because the development and better utilization novel machine learning for classifying non-cancer deaths in cancer patients is expected to reduce the morbidity and mortality in these patients. Thus, the proposed research is relevant to the part of the NIH's mission that pertains to developing fundamental knowledge that will help to lengthen human lives and reduce the burdens of human illness.

由于医疗保健的风险很高，主要障碍是对错误的容忍度极低 医疗保健实践，这要求任何建模都具有极高的敏感性和特异性。然而， 几乎所有机器学习 (ML) 模型都专注于提高准确性。目前尚无法同时达到这两个目标 使用医疗保健数据具有极高的敏感性和特异性。单独的筛选和验证性机器学习 建议使用工具来实现非常高的灵敏度和特异性。此外，许多机器学习算法都受到影响 由于缺乏明确的解释，例如深度学习和神经网络，并且不太可能满足 公平标准。癌症是美国第二大死因 癌症幸存者人数 持续增长；不幸的是，癌症患者的非癌症死亡人数也在增加。然而， 几乎所有的组学和大群体研究都集中于二元结果（癌症死亡或复发）。 因此，迫切需要更好地了解和减少癌症患者的非癌症死亡， 使用“组学和人口数据”。为了解决这些问题，该项目将开展筛选和 验证性机器学习可模拟乳腺癌、结直肠癌、前列腺癌和肺癌的癌症和非癌症死亡 使用“组学数据”和电子健康记录（EHR）的患者。拟议的研究将导致 对机器学习工具、工作流程和方法做出了根本性贡献，以新颖地利用“组学”和 EHR 数据 用于癌症护理。及时满足了精准减少非癌症死亡的迫切需求。这个项目 还独特地解决了变革性数据科学研究主题。跨学科 合作计划中概述的该项目的合作将为创意提供多样化的基础 问题解决和验证。该提案具有 3 个更广泛的影响：1）开发的新颖的 ML 算法和技术将使医生能够更准确地预测和治疗癌症 根据患者多类别死亡的风险。 2）该研究计划将支持和培育 本科生和研究生研究人员。 3) 拟议的研究计划将支持高中和 本科生在进行研究和了解 ML 的实用性方面。 相关性（参见说明）： 拟议的研究与公共卫生相关，因为新颖的开发和更好的利用 用于对癌症患者的非癌症死亡进行分类的机器学习有望降低发病率 和这些患者的死亡率。因此，拟议的研究与 NIH 的部分使命相关 这涉及发展有助于延长人类寿命和减少死亡的基础知识 人类疾病的负担。