Intelligent Healthcare Systems for Large-scale Populations

This project will be dedicated to developing and applying core AI technologies for general health data sciences. So far, I have developed three workable AI models for detecting infant stroke on a small dataset of accelerometer data collected by the Institute of Neuroscience, Newcastle University. Additionally, I am an expert in deep learning and am experienced at dealing with large-scale complex multi-modal data. My expertise in Zero-shot Learning focuses on addressing model interpretation and data insufficiency problems. I have additional strength in programming and software engineering. The project can be summarised into two main stages. The first stage focuses on proof-of-concept studies on two historical datasets, UK Biobank and NE 85+. UK Biobank has provided baseline measurements (such as the eye measures and saliva samples). In addition to the baseline assessment, 100,000 UK Biobank participants have worn a 24-hour activity monitor for a week, 20,000 of whom have undertaken repeated measures. A programme of online questionnaires is being rolled out (diet, cognitive function, work history and digestive health) and UK Biobank has embarked on a major study to scan (image) 100,000 participants (brain, heart, abdomen, bones & carotid artery). UK Biobank is linking to a wide range of electronic health records (cancer, death, hospital episodes, general practice), and is developing algorithms to accurately identify diseases and their subsets. Blood biochemistry is being analysed (such as hormones & cholesterol). Genotyping has been undertaken on all 500,000 participants and these data are being used in health research. In NE 85+, a total of 484 participants aged 87-89 years recruited to the study completed a purpose-designed physical activity questionnaire (PAQ), which categorised participants as mildly active, moderately active and very active. Out of them, 337 participants wore a triaxial accelerometer on the right wrist over a 5-7-day period to obtain objective measures. Data from subjective and objective measurement methods were compared.

The first stage of the project utilising the NE 85+ data aims to integrate these complex data, e.g. MRI, accelerometer, and electronic health records. The project development will follow a simple-to-complex logistic. Initially, only one disease and one factor will be considered. Subsequently, multiple factors will be simultaneously considered. The model will then be progressively upgraded to take into account different health data sources and the correlations between different diseases. At this stage, we focus on disease diagnosis and alert, i.e. prediction of the risks of diseases based on observed factors. After stable performance has been achieved, the model will be focused on the rationale study. For example, what lifestyle or other factors can result in a high risk of heart disease? What is the attribute that we can change to reduce such a risk? Beside these rationale studies and healthcare feedback, visualisation techniques can provide more qualitative results that can help medical experts to discover new knowledge.

During the second stage, stable AI models can be packaged into apps, with objective of encouraging more participants to engage in the study. Through smartphone or wearable sensors, participants can get access to direct healthcare from the cloud-based AI server. In turn, the collected data will be used to upgrade the model, validate previous studies, and large-scale cohort clinical study. More details can be found in the technical summary.

The first stage focuses on proof-of-concept studies. After getting access to UK Biobank and NE 85+, the inital step is Data Preprocessing, i.e. integration of complex data, such as MRI, accelerometer, and electronic health records. My previous AI model for infant stroke detection can be employed to explore some preliminary experiments, e.g. associations between human activities and specific diseases. Following this, utilising a multi-modal learning framework, a comprehensive model will be proposed to study the correlations between different health data sources. For example, what MRI pattern can correspond to a disease and how would that relate to the accelerometer data for human activity. Furthermore, attribute learning techniques can provide health feedback, e.g. how to implement a habit that can significantly reduce the risk of a disease. Finally, the model will be upgraded into an inference model and link the predictions to attributes, in a manner that is intuitive to understand.To further improve the interpretations, high-level visualisation technology, e.g. tSNE, Confusion Matrix, will provide furhter guidance about rationales behind the AI model.

During the second stage, proof-of-concept studies will continue to the end of the project. Meanwhile, I plan to implement stable models on apps and smart sensors to encourage more participants to take parts in data collection. For example, stable models can be packaged into apps that can utilise mobile phone-embedded accelerometers as an approximate measurement for data collection and large-scale cohort clinical study. The expected outcomes of the project will be a series of high-quality publications both in AI and health data science fields. In addition, continued grant proposals will be submitted for the continued study of large-scale health data collection using the delivered system.