Authors (including presenting author) :
LAM DCC (1), A.Y.Y. Ng (1) KM Cheung (2), P.Y.M. Woo (3, 5), K. Y. Chan (4) D.T.M Chan (5)
Affiliation :
(1) Mechanical engineering, HKUST (2) Clinical Oncology, Queen Elizabeth Hospital, (3) Medicine, The Chinese University of Hong Kong (4) Neurosugery, Kwong Wah hospital, (5) Neurosugery, Prince of Wales Hospital
Introduction :
Health care data from blood tests to images records to diagnostic notes and medication records are collected at primary health care and in hospitals in the territory. Mining of health care data and refining the data into information helps determine if the subject has chronic diseases. Complicated diseases such as strokes, cancers and heart attacks are hidden inside the collected evidence. Smart clinics and smart hospitals armed with rich data and advanced AI tools (NCD vectoring) can potentially identify high risks NCD groups from the DM/HT/HL groups for prioritized imaging and preventive intervention before the NCDs such as cancers reach the late stage and stroke, and heart diseases become acute. The progress in the development of pathways to transform primary health care clinics and hospitals into smart heath care centers are outlined.
Objectives :
Primary health care centers provide community health care for DM/HT/HL patients, and the upgrading of primary health care centers opened clinics to offer laboratory tests and imaging that had been available mainly in hospitals. Health and laboratory data collected at primary health care can reveal NCDs hidden in DM/HT/HL population if tools are available to boost the sensitivity and specificity to reveal the hidden NCDs. The objective of the project is to develop NCD detection tools for multiple risk cohorts with varied risk profiles and pre-existing conditions for a selection of NCDs. The tools are integrated to form NCD vectors to improve early detection of hidden NCDs, smart triage and high feedback monitoring of NCDs in clinics and hospitals.
Methodology :
Establishment of HADCL made territory-wide data accessible for mining and testing of the concept. Selected NCD vectoring tools are developed covering strokes, cancers and heart attacks. 20 years of Hong Kong data were collected into NCD data sets, cleaned and transformed the multi-dimensional evidence labeled streams for machine learning. Machine learning was used to align, integrate and scale the high dimensional streams into risk scores. Alignment with clinical notes and correlations reported in the literature were used to benchmark the predictions and to explain the AI findings to ensure that the AI results are supported by a clear clinical evidential foundation.
Result & Outcome :
NCD vectors for strokes, cancers and heart attacks are developed. The utilities of NCD vectors for the detection of intracranial aneurysms and for assessing the rupture risk are presented. 2M patient records were collected from database at HADCL. 1M records containing ischemic strokes, AMI, CHD and cancers were excluded. 1M patient records were left after exclusions. 393 subjects with ICA and 4,096 subjects with brain images without ICA were used as control. Laboratory data including CBC, RFT and LFT and brain images were collected for the cohort for data treatment and machine learning training. The NCD tools were tested using a cohort from KWH. The NCD vector showed that the rupture can be accurately identified using laboratory data yielding a PPV of 0.78 and NPV of 0.9. The inclusion of aneurysm images boosted the PPV to 0.86 and NPV to 0.96. The case study showed that laboratory blood test including CBC, RFT and LFT can accurately identify ICA early and predict its rupture risk in primary health care.
