| Summary: | Post-stroke adverse mental outcomes (PSAMO) such as depression and anxiety are
common comorbidities of stroke. A study has shown that about 30% of stroke survivors have
depression and about 20% of stroke survivors developed anxiety at some point after stroke.
Stroke survivors with such adverse mental outcomes have been shown to have poorer health
outcomes such as higher mortality and greater functional disability. This study aims to use
Artificial Intelligence (AI) such as Machine Learning (ML) algorithm and Deep Learning
(DL) architectures to develop a predictive model that predicts the risk of PSAMO.
A retrospective study of 1,780 patients with stroke episodes is collected for this study.
Features collected include patients’ demographic and sociological data, quality of life scores,
stroke-related information, medical history and more. A series of ML algorithms are trained
and tuned using Bayesian optimization used to select the optimal hyperparameters.
eXplainable AI (XAI) methods such as SHapley Additive exPlanations (SHAP) is then used
to interpret the Machine Learning (ML) model. DL models are developed using various
architectures such as Multi-layer Perceptron (MLP) and Long-short Term Memory (LSTM).
Sequential data such as time-series of daily lab results is also trained by the DL model.
The best performing ML model, using Gradient Boosted Trees, achieved 74.7%
accuracy. Using DL architecture, using a combination of MLP and LSTM, the accuracy
improved to 77.8%. The application of DL algorithms, being able to capture complex non
linear relationships and having the flexibility of training a model using different data types,
boosted the accuracy in the prediction of PSAMO.
We conclude the study proposing that using ML algorithms and DL architectures are
able to predict the risk of PSAMO with good accuracy. This study is also the first to use DL
architectures for the prediction of PSAMO. In addition, the adoption of DL architectures
allows the model to ingest different data types for prediction, which improved prediction.
|
|---|