SCH: Multidimensional Microfluidic Salivary Sensor with Adversarial Knowledge Distillation for Point-of-Care Assessment of Periodontitis and Comorbidities

本研究的目标是开发一种传感器设备原型，以快速测量一系列不同的 唾液生物标志物作为新型机器学习（ML）方法的输入，可以预测牙周炎， 监测牙周进展。我们的长期目标是开发一种快速、用户友好、低成本的定点护理（POC）设备，用于牙医办公室或家中，快速集成和分析数据 以支持患者管理。它解决了数据科学，计算生物学， NIDCR的生物信息学计划在整合和分析大量和多样化的数据，以更好地 了解牙科，口腔和颅面生物学和疾病。 据CDC称，近50%的成年人患有某种形式的牙周病。牙周炎是无声的 进展和患者往往寻求专业护理，只有在先进的阶段，先进的，痛苦的 并且需要昂贵的程序来控制疾病或替换掉的牙齿。牙周病的早期诊断 需要个体患者水平的疾病，并且越来越多的人意识到， 在预测个人患病风险方面很有价值。我们假设可以建立预测模型 基于对唾液中大量牙周炎相关生物标志物的测量;传感器设备 集成多传感器模态和机器学习（ML）模型的技术将推进临床目标 牙周炎的早期诊断，使早期的临床干预。因此，我们将开发和应用 用于检测与不同阶段相关的唾液分析物浓度的三种不同的传感器模式 牙周病的进展，即，炎症、软组织破坏或骨破坏（目的1）。数据从 传感器输出和临床检查都将用于通过一种新的多模态 对抗性知识蒸馏ML框架，它促进了准确的早期预测， 纵向数据表示（目标2）。多传感器模态和ML模型将被嵌入 在单个微流体装置中，将诸如采样、检测和数据分析的步骤合并为微流体装置， 集成的芯片实验室，并允许传感器数据预处理，以仅传输可操作的 将信息提供给外部平台，以保护用户的隐私（目标3）。这样的设备预计将提供 进行不显眼、准确和频繁的基于唾液的自我监测，并提供详细的医疗数据， 支持临床决策。它将是未来个性化医疗的有效工具， 改善患者口腔健康。

The goal of this research is to develop a sensor device prototype to rapidly measure an array of diverse salivary biomarkers as input for novel machine learning (ML) methods that can predict periodontitis and monitor periodontal progression. Our long-term goal is to develop a rapid, user-friendly, and low-cost pointof-care (POC) device, for use in either a dentist’s office or at home, that rapidly integrates and analyzes data to support patient management. It addresses the priority area of the Data Science, Computational Biology, and Bioinformatics Program of NIDCR in integrating and analyzing high-volume and diverse data to better understand dental, oral, and craniofacial biology and diseases. According to the CDC, nearly 50% adults have some form of periodontal disease. Perioodontitis is silently progressive and patients often seek professional care only in an advanced stage where advanced, painful and costly procedures are needed to control disease or replace lost teeth. Early detection of periodontal disease at an individual patient level is required and there is growing awareness that multiple biomarkers are valued in predicting risk of disease in individuals. We hypothesize that predictive models can be established based on the measurements of a large set of periodontitis-associated biomarkers in saliva; a sensor device that integrates multi-sensor modalities and the machine learning (ML) models will advance the clinical goal of early diagnosis of periodontitis to enable earlier clinical interventions. Thus, we will develop and apply three distinctive sensor modalities for detecting concentrations of salivary analytes relevant to various stages of periodontal progression, i.e., inflammation, soft tissue destruction or bone destruction (Aim 1). Data from both sensor outputs and clinical examination will be used to train ML models via a novel multi-modal adversarial knowledge distillation ML framework, which promotes accurate early prediction with partial longitudinal data representations (Aim 2). The multi-sensor modalities and the ML models will be embedded in a single microfluidic device, incorporating steps such as sampling, detection, and data analysis as an integrated lab-on-a-chip, and permitting the sensor data preprocessed to transmit only the actionable information to the outside platform to protect the user's privacy (Aim 3). Such a device is anticipated to offer for unobtrusive, accurate, and frequent saliva-based self-monitoring, and provide detailed medical data to support clinical decisions. It will be an effective tool for future personalized medicine and dramatically improve patients' oral health.