Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage News Making Machine Learning Adoptable for Clinicians

Making Machine Learning Adoptable for Clinicians

Leia em Português

Dr. Alexander Scarlat explains the core tenants of machine learning in his 12-part series "Machine Learning Primer for Clinicians". The texts cover defining aspects of machine learning in the first seven parts, followed by examples that communicate aspects of measuring the performance of machine learning models. The series uses animated charts in place of the math to help readers understand the machine learning concepts.

The series starts by focusing on the deficiencies in traditional rule-based algorithms while at the same time explaining that machine learning can overcome those same challenges. Challenges include multi-dimensional problems and human limitations to statistical program rules.

Scarlat effectively builds a machine learning concept taxonomy for the reader. At a high level, the series explains supervised, unsupervised, and other forms of machine learning. Supervised training includes regression and classification. Unsupervised includes clustering and anomaly detection. Other forms of machine learning include ensemble methods and generative models. The entire article series follows a similar approach to the progressive elaboration of a taxonomy of defining machine learning concepts.

In part four, Scarlat explains the limitations and requirements for the pre-processing of data used in machine learning. Pre-processing includes flattening relational data, adding values where they are missing, converting text into either sparse arrays or vectors, using binary values in separate vector dimensions for categories. Additionally, Scarlat covers the need for normalization of values to prevent a false sense of significance in magnitude and the curse of dimensionality.

Part five of the article series describes how machine learning algorithms learn, including the use of weights, metrics for performance, loss (or cost) function, using an optimizer, forward propagation, and backpropagation. Beyond that, in part seven, additional information is found about the workflow of creating a machine learning model, including underfitting, overfitting, training, validation, testing, learning rate, data augmentation, using regularizers, and using dropout.

Scarlat completes the texts by using four different real-world problems and example solutions to explain evaluating the performance of machine learning models. His article titles include:

The examples communicate the performance metrics including precision, accuracy, recall, the area under the curve, and F1 score. Additionally, the last model shows how existing machine learning models can be used to transfer learning and minimize additional model training.

Additional topics that enable adoption of machine learning in healthcare include: identifying the processes to use machine learning in, integration with existing electronic health record (EHR) systems, and capturing of care outcomes. Google AI applies machine learning to EHR patient records with a focus on scalability, accuracy, and interpretability.

Rate this Article