Machine learning in medicine using JavaScript: building web apps using TensorFlow.js for interpreting biomedical datasets

Abstract

Introduction: Contributions to medicine may come from different areas; and most areas are full of researchers wanting to support. Physists may help with theory, such as for nuclear medicine. Engineers with machineries, such as dialysis machine. Mathematicians with models, such as pharmacokinetics. And computer scientists with codes such as bioinformatics. Method: We have used TensorFlow.js for modeling using neural networks biomedical datasets from Kaggle. We have modeled three datasets: diabetes detection, surgery complications, and heart failure. We have used Angular coded in TypeScript for the implementation of the models. Using TensorFlow.js, we have built Multilayer Perceptrons (MPLs) for modelling our datasets. We have employed the training and the validation curves to make sure the model learnt, and we have used accuracy as a measure of goodness of each model. Results and discussion: We have built a couple of examples using TensorFlow.js as machine learning platform. Even though python and R are dominant at the moment, JavaScript and derivatives are growing fast, offering basically the same performance, and some extra features associated with JavaScript. Kaggle, the public platform from where we downloaded our datasets, offers a huge amount of datasets for biomedical cases, thus, the reader can easily test what we have discussed, using the same codes, with minor chances, on any case they may be interested in. We were able to find 92% of accuracy for diabetes detection, 100% for surgery complications, and 70% for heart failure. The possibilities are unlimited, and we believe that it is a nice option for researchers aiming at web applications, especially, focused on medicine.

Robodoc: a conversational-AI based app for medical conversations

“…The former was trained on numerical medical datasets, whereas the latter was trained on medical image datasets. The numerical model was trained to detect diabetes Pires (2023b). The image model was trained to detect anomalies in OCT images and pneumonia from X-ray images Kermany et al (2018).…”

“…What is needed are their datasets, and main instructions they have followed. The number-based model is from a previous work of mine Pires (2023b).…”

“…Table 2 illustrates that the text-triggered path behaves as expected. What I need to consider in the future is how this path will behavior when more models are added, such as the ones from Pires (2023b). It is natural that it is considered how the system will be scaled up, how the system will behavior as we add new models, which will add new capabilities.…”

“…I am going to focus on a shallow neural network, with no transfer learning, and a small number of layers and neurons. The model I am going to use, and possible variations, is from a previous work of mine Pires (2023b).…”