Historical data can help to improve present performance, and when such data is made available in real time, it becomes even more valuable. However, detecting intricate patterns from data is tough for traditional statistical models. Machine learning with neural networks, however, has shown remarkable results in such tasks – on structured as well as unstructured data.
With my recent foray into machine learning based on Prof Andrew Ng’s excellent courses (https://www.coursera.org/learn/machine-learning & https://www.coursera.org/specializations/deep-learning), and over 15 years data available from the acute stroke care program at Baby Memorial Hospital, the circumstances were ripe to attempt building a neural network – for predicting potential delays based on parameters recorded at the time of admission.
The full technical details of how this was designed and implemented are available at https://mcareapps.com/docs/design_neural_network_stroke_pathway.pdf
The parameters that influenced treatment time were identified as the duration of symptoms (onset to arrival time), age, time of day and severity of stroke. However, taken individually, the effects were very weak. Using these four parameters from over 200 patients, a 3-layered neural network implemented using the TensorFlow framework. After training, the network achieved an accuracy of ~ 60% in predicting delays.
This was then validated on an independent set of data.
The next step is showing a real-time notification: “Predicted DNT > 45 min!” (Door to Needle Time, stands for the time interval from hospital arrival to the initiation of clot-removal therapy) when a new patient from BMH with potential delay is entered in the MCare Acute Stroke app.

The accuracy of prediction should gradually go up, as the neural network is trained with more data.
In an emergency like stroke, every minute matters.
Machine learning AI trained on historical data has the potential to improve care by informing the team in advance about possible delays.
[Technical side note: Getting the neural network working on the Node.js app backend was done by implementing the forward part of the network, along with the trained parameters from TensorFlow, in Javascript.]