Health care predictive analytics using artificial intelligence techniques
- Publication Type:
- Thesis
- Issue Date:
- 2018
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
In recent years, advances in Artificial Intelligence (AI) are opening the door for intelligent health care data prediction and decision making. Machine learning, as an increasingly popular approach to AI, has been widely used to learn directly from data, adapt independently, and produce predictive outcomes, which support doctors when encountering complex health care predictive analytics. However, traditional machine learning methods are not always perfectly working in the health field, intrinsically due to little consideration for characteristic problems within health care data. For example, the small sample size problem is common due to complex data collection procedures and privacy concerns. Missing data is also widely encountered since most data are collected as a second-product of patient-care activities instead of following systematic research protocols. The class imbalance is another inevitable problem in the medical data as the normal class always predominates over the disease class. To solve aforementioned issues in health care predictive analytics, this study stands on the principles of machine learning and transfer learning to develop five advanced prediction models.
The first model is an output-based transfer least squares support vector machines (LS-SVMs) model which can leverage knowledge from the existing prediction model or on-line tool to facilitate the learning process on the current domain of interest with insufficient data. This model overcomes the small sample size problem and improves the health care data prediction by learning knowledge from the other domain.
The second model is a novel additive LS-SVMs model which can make predictions simultaneously considering the influences on the classification error caused by missing features in a dataset. This model can generate valuable explanations regarding the influence levels of missing features for health professionals to improve the future data collection process.
The third model is a transfer-based additive LS-SVMs model which can deal with missing data from a transfer learning perspective. It can leverage the model knowledge learned from the complete portion of the dataset to help the learning process on the whole dataset with missing data. The proposed model can provide supplementary information for health professionals to improve the data quality via data cleaning.
The forth model is a deep transfer additive LS-SVMs model called DTA-LS-SVMs and its imbalanced version called iDTA-LS-SVMs to enhance the prediction performance on the balanced and imblanced datasets. Inspired by the stacked architecture and transfer learning mechanism, the model stacks multiple additive LS-SVMs based modules layer-by-layer and embeds model transfer between adjacent modules to guarantee their consistency.
The fifth model is a deep cross-output transfer LS-SVMs model called DCOT-LS-SVMs and its imbalanced version called IDCOT-LS-SVMs to improve the prediction performance on the balanced and imbalanced datasets. The cross-output transfer is used to transfer the predictive outcome from the previous module to the adjacent higher layer to achieve a better learning. Moreover, modules’ parameters can be randomly assigned in the proposed model which significantly reduces the time for model selection.
The proposed models are verified using experiments on the public UCI datasets. Moreover, case studies are conducted to validate and integrate the proposed models with real world applications, including bladder cancer prognosis, prostate cancer diagnosis, and predictions of elderly quality of life (QOL). The results have demonstrated that the models in this study can enhance the prediction performance while taking the characteristic problems within health care data into account, thus exhibiting promising potential for use in different health applications in future.
Please use this identifier to cite or link to this item: