Deep learning will revolutionize healthcare - Dr. Sachin Gavali

What is deep learning?

Ever since Von Neumann and Alan Turing conceived the idea of modern computers, scientists have dreamt about creating computer systems that can mimic human intelligence.

Initial efforts at developing intelligent programs were geared towards building increasingly sophisticated rules to handle the logic that was thought to govern intelligence. Eventually, researchers realized the complexity of building these rules was intractable, and a better approach was needed. Hence, instead of writing these rules manually, researchers created programs to generate these rules from existing data without any human intervention.

The study of building programs that can learn and extract information from data is known as machine learning.

A subfield of machine learning and a stepping stone to the coveted dream of artificial intelligence is the study of Deep Neural Networks (DNN). The fundamental unit of a DNN is the perceptron, a coarse computational representation of a biological neuron and thus aims to mimic the human brain.

Though the study of deep neural networks has its roots in the 1950s, their true potential was demonstrated by Alex Krizhevsky, who introduced the first deep learning model - AlexNet. The model showed that it was possible to use DNNs to recognize and understand the content of images without any human intervention. Since then, the field of DNN research has exploded tremendously and has been progressing in three distinct directions - speech, vision, and natural language, each roughly trying to address an individual aspect of human cognition.

Applications of deep learning in healthcare

1. Drug discovery and Precision Medicine

Various well-known pharmaceutical companies such as Roche, Bayer, and Pfizer have invested significantly in deep learning systems to aid numerous drug discovery and development stages. One of the earliest applications of DNNs has been to develop novel drugs and therapeutics.

As an extension of this application, DNNs have also proven to be helpful in precision medicine. This is particularly evident by the recent introduction of an extremely powerful deep learning system named AlphaFold2 created by the researchers at DeepMind.

Now, researchers can use newer systems to simulate interactions of proteins without performing costly and time-consuming experiments.

AlphaFold2 can accurately predict three-dimensional structures of proteins with an error margin of ~1.6 Angstroms. This has effectively solved a 50-year-old problem in molecular biology and revolutionized the field of novel drug development.

2. Augmenting clinical decision making

Traditionally, healthcare practitioners have relied on their domain knowledge and intuition decision-making in diagnosis and treatment modalities.

In recent years, computer-assisted diagnosis software has seen an increased adoption to provide clinicians with a second opinion. However, these software systems have proven inadequate more often than not.

Also, in recent years, systems such as the IBM Watson, which make use of massive datasets ranging from biomedical literature, ontologies, genomic and pharmaceutical studies, have shown promising results.

In addition to integrated approaches that handle many aspects of clinical decision-making, simpler targeted systems have also proven valuable. For example, many hospitals maintain an electronic record of patients. These records contain everything regarding diagnosis, treatments, and follow-ups about a patient. Usually, these records are made by doctors while handling the patients and thus are not in a well-structured format. Deep learning systems based on the principles of natural language processing have proven to be immensely useful in extracting actionable information from these records.

3. Forecasting the population health dynamics

Predicting the trajectory of large-scale health disorders such as epidemics has always been a challenge for various government and non-government agencies. Traditionally researchers employed classical statistical methods such as ordinary least square regression to study and analyze the factors that govern the trajectory of epidemics.

However, contrary to the factors driving health conditions at the individual level, those at the population level are not constant. They are affected by various continuously changing factors across both time and geography. This presents a significant challenge in the way one can use data to build predictive models for epidemics.

Traditional statistical methods require data in a well-defined format collected according to some predefined protocols.

However, a recent class of deep learning methods based on the principles of graph theory has shown different results. It has shown that one can use the data not collected according to any predefined protocols and yet build predictive models and understand the trajectory of large-scale epidemics.

Challenges

Despite significant advancements in deep learning methods, numerous challenges have prevented their widespread adoption in healthcare. The most prominent one is the apparent black-box nature of these models.

For clinicians to have confidence in deep learning systems, they also need to explain the predictions. However, providing explanations for predictions is a challenge. Creating explainable models leads to simpler models with lower predictive performance. Thus explainability is inversely proportional to predictive performance.

However, in recent years a new class of model explainability techniques is being developed - the most promising of them being Shapley Additive Explanations (SHAP).

SHAP is inspired by the principles of game theory, where the task of model prediction is considered a game played by the factors that determine the outcome of the task. The explanations of the predictions are then derived by quantifying the relative contribution of these factors to the outcome.

In addition to explainability, the lack of adequate data has also been one of the significant challenges to creating deep learning models for healthcare applications. Most of the data available in the healthcare domain is sparse and locked behind proprietary data usage agreements.

Though some organizations have started adopting the FAIR (Findability, Accessibility, Interoperability, and Reuse) principles to share their data, there is still a need to change the culture surrounding data use and access for research in healthcare domain.

About the author

Sachin Gavali is pursuing his Ph.D. in Bioinformatics Data Science at the University of Delaware (USA) under the mentorship of Dr. Cathy Wu.

Currently, his research is focused on developing novel deep learning algorithms for problems in the biomedical domain.

He also works on the development of high-performance machine learning systems with a focus on large-scale knowledge discovery from existing biomedical literature.

Before joining Ph.D., Sachin completed his bachelor's degree in Dentistry from Terna Dental College, Mumbai.