NeuReader: Eye Tracking Enabled Explainable-AI for Empowering Resource Scarce Neurological Healthcare in Pakistan

神经疾病给巴基斯坦的医疗系统带来了巨大的负担，在巴基斯坦，只有少数训练有素的神经科医生可以为2.31亿人口服务。巴基斯坦60%以上的人口居住在农村地区，通过初级医生或护理人员运营的基本卫生单位(BHU)和农村卫生中心(RHC)提供医疗保健。这些设施没有人力或资源来提供任何神经护理。NeuReader项目的目的是开发一种系统，帮助改善巴基斯坦患者的神经护理。将使用脑电(EEG)监测神经健康状况，脑电通过放置在头皮上的电极测量大脑的电活动。NeuReader将利用人工智能和自然语言处理来读取EEG数据，诊断神经疾病，并提供解释问题的报告。没有任何解释的简单诊断决定没有太大帮助，因为初级医务人员在母婴医院或皇家医院进行测试也需要帮助提高患者对其健康状况的认识。更好地了解诊断还将使患者及其家人能够在知情的情况下决定前往城市地区的医院寻求进一步的帮助，因为这需要大量的旅行和住宿费用。NeuReader的可解释性功能还将帮助神经科医生(远程连接到系统)根据患者的病情严重程度对患者进行优先排序。建立这样一个系统需要克服几个技术挑战。人工智能系统的性能在很大程度上取决于可用于训练它们的数据的数量和质量。训练将使用两种类型的数据：(1)带有医生报告的EEG记录，并用文字进行总结(2)EEG记录中发现的异常位置。医生报告将由神经科医生在数据收集期间撰写。在脑电记录中标记异常位置既耗时又费力。EEG记录可以包括几分钟/几小时的数据，异常仅持续几秒，并且分布在记录中的不同位置。为了避免花费数百个小时标记脑电记录，眼球跟踪将被用来记录神经科医生在日常工作中检查脑电数据时计算机屏幕上凝视模式的位置。然后，这些眼睛凝视模式将被用来迅速生成事件的标签，从而节省了几个小时的宝贵神经科医生的时间。录制的标签将被用来训练人工智能算法，该算法可以自动发现感兴趣的事件，然后可以用来生成文本报告，供母婴健康中心的初级医生和护理人员使用，以帮助疑似患有神经疾病的患者。将有大量时间专门用于实地研究，旨在评估将成为该系统最终用户的患者和医生的需求。这些实地研究的学习成果将被纳入最终设计，以最大限度地发挥地面影响。

Neurological disorders place a significant burden on the healthcare system of Pakistan where only a handful of trained neurologists are available to serve a population of 231 million people. More than 60% of Pakistan's population resides in rural areas where healthcare is provided via basic health units (BHU) and rural health centres (RHC) which are run by junior doctors or nursing staff. These facilities do not have the manpower or resources to provide any neurological care. The aim of project NeuReader is to develop a system that helps improve the provision of neurological care to patients in Pakistan. Neurological health will be monitored using electroencephalograms (EEG) which measure the brain's electrical activity via electrodes placed on the scalp. NeuReader will leverage Artificial Intelligence and Natural Language Processing to read EEG data, diagnose neurological disorders, and provide a report explaining the problem. Simple diagnostic decisions with no explanation are not very helpful as junior medical staff administering the tests at BHUs or RHCs would also need help to increase patient awareness about their health condition. A better understanding of the diagnosis will also allow patients and their families make informed decisions about travelling to hospitals in urban areas to seek further assistance since that entails substantial travel and lodging expenses. The explainability features of NeuReader will also help neurologists (connected remotely to the system) prioritise patients based on the gravity of their conditions. Building such a system requires overcoming several technical challenges. The performance of AI systems is highly dependent on the amount and quality of data available for training them. Two types of data will be used for training: (1) EEG recordings with doctor's reports summarising them in words (2) Locations of abnormalities spotted within an EEG recording. Doctor reports will be written by neurologists during data collection. Labelling of locations of abnormalities within EEG recordings is time consuming and laborious. An EEG recording may consist of several minutes/ hours of data with abnormalities lasting only a few seconds and spread out across different locations within the recording. To avoid investing hundreds of hours labelling EEG records, eye tracking will be used to record locations of neurologist gaze patterns on a computer screen as they examine EEG data in their routine practice. These eye gaze patterns will then be used to promptly generate labels of events saving hours of highly valuable neurologist time. The recorded labels will be used to train AI algorithms that can automatically spot events of interest which can then be used to generate a text report that can be used by junior doctors and nursing staff at BHUs or RHCs to assist patients suspected of suffering from neurological disorders. A significant time will be dedicated to field studies designed to assess the needs of patients and doctors who will be the end users of this systems. The learn outcomes of these field studies will be incorporated into the final design to maximise on ground impact.