Alzheimer’s disease is an irreparable, degenerative disease with ongoing loss of functions of the brain. Currently, there is no medicine or treatment present to stop or slow down the progression. The identification of different stages for diagnosis require a combination of clinical data, complex cognitive tests, radiology imaging, demographic information, time and highly skilled physicians. Recent machine learning techniques can help to provide a process to extract insights and improve the quality of life for the patients and assist the physicians.
In this project, various machine learning techniques such as feature selection, feature engineering, dealing with imbalanced data, imputation of missing values and standardization are applied. Multiple algorithms are also compared before performing random grid search to tune hyper-parameters for classifiers and developing an ensemble learner to classify three clinical stages (normal, mild cognitive impairment and dementia). Cognitive tests, magnetic resonance imaging of left hippocampus and cortical thickness of right entorhinal are discovered to be important features for prediction. This finding is similar to that reported in a number of studies. The model can equally distinguish between a class and other classes with an average area under the receiver operating characteristics score of 0.83. This is within the range of the evaluation metric of
existing state-of-the-art models.
A web-based application is developed and deployed to the cloud to address the gap for the user to benefit from the developed model. This result in an end-to-end pipeline that will empower the user with a practical application and contribute to the active research in the area.
The code is available at : https://github.com/piushvaish/ad_ncirl
The proposed approach can be improved in the future in the following ways:
• The resulting predictions just indicate the three stages of Alzheimer’s disease. This can be extended to more granular stage
• The predictions can be more probabilistic in nature to include the likelihood of diagnosis and present the next course of action for treatment. • The results of the model can be logged and monitored to ﬁnd the performance of the machine learning model over time.
• The web-based application is simple and provides a proof of concept. The features and model used to create can be improved further along with enhanced user design and experience.