Last year, CB Insights published a list of 106 startups that are using artificial intelligence (AI) and machine learning (ML) in various areas of the health industry. These startups covered a phenomenal range of activity – from patient data and risk analytics (identifying high-risk individuals for targeted intervention) to medical imaging, wearables and virtual assistants.
AI and the health industry have been on a long journey together to get to this point, but we are still some way from the potential of these technologies being realized. There are challenges standing in the way of greater adoption of AI in healthcare – particularly when it comes to data quality – but the industry is working hard to get to grips with them. As we’ll see, there is certainly much hope to be found in the growing role of Chief Data Officers in healthcare organizations.
There is no consensus in the industry and academia about the definition of AI. A recent Rock Health report defines AI as human-like capabilities of specific mathematical algorithms processed by computers. The ability to learn from ‘errors’ is a key attribute of these algorithms.
In 2010, there was groundbreaking progress in the field of AI owing to the advent of multi-layer neural networks that could be trained deeper and faster thanks to the explosion of computing power. The availability of large labelled data sets (such as images, web queries and social networks) boosted the training testbeds that provided much-needed momentum to move towards deep learning (DL) on deep neural networks (DNN), including an architecture termed convolutional neural networks (CNNs). The latter powered major breakthroughs in image classification, meaning that image analysis emerged as an exciting new application area for AI and ML algorithms.
AI and ML in healthcare
Technology companies like Google, Facebook and IBM have been working on algorithms for face and image analysis for a long time and with some success. That knowledge and technology is now being applied directly to image analysis in health care – such as the retinal images used for identifying diabetic retinopathy or dermatological images used for determining melanoma or skin cancer. Medical professionals have studied and labelled millions of images with their own notes, which can now be used to train DL algorithms to detect the early onset of melanoma or the presence of diabetes.
Applying AI or ML in a general clinical context, however, remains problematic. The availability of large datasets in a general clinical setting (outside of imaging systems) is still questionable and healthcare data is usually episodic. Just 5% of healthcare users consume 50% of healthcare resources, meaning that healthcare data is sparser than in other domains. Data collected from electronic health record (EHR) systems, therefore, reveals a lot of data about a few patients and very little to nothing about the rest of the population. This introduces another challenge.
Health is a complex function of many variables: medical, environmental, genetic and social.
In today’s healthcare systems the primary digital data source is medical, which is stored in the EHR. These systems were designed many decades ago for billing purposes and still not compatible with the growing needs to acquire and store new data sources. In the recent past some progress has been made in terms of capturing social determinants of health like homelessness, unemployment, food and transportation needs into the medical record, but a lot more is required.
This is linked to the limits of using AI and ML with genetic data to support clinical decision-making. An algorithm that works in development under test conditions might not be useful in the real world. Diseases could be related to multiple mutations or exposure to environmental toxicity, which is not currently measured for patients. Therefore, clinical trial is necessary to prove that they can help improve decision-making in practice.
Chief Data Officers to the rescue?
Based on my conversation with many Chief Data Officers (CDOs) at the IBM CDO summit 2017 and also from my personal experience as a CDO, I can say that the position is evolving. What used to be a traditional data governance, data management and regulatory compliance role is now one that drives more decisions within the organization. Although many companies are still unclear on how to use these roles for building a more coherent data-driven strategy, a Gartner survey from 2017 shows that CDOs have larger areas of focus as well as bigger budgets than previous years.
One of the biggest challenges facing data leaders in healthcare organizations is resolving data quality issues. Since the correction of data acquisition errors often leads to EHR data entry workflow re-engineering, this is often difficult politically. Data privacy regulations create significant barriers for seamless data interoperability between organizational boundaries, as pointed out by the US Department of Health and Human Services’ 2017 study on the impact of AI on healthcare in the near and longer term. The report also underscored that training data in healthcare may not closely match what will be encountered in real-life application. Diseases can change and new diseases can appear, which makes any AI/ML system that is launched today less and less effective over time. Algorithms must be assessed continuously to understand how to respond to such changes.
There are many reasons to be hopeful out there, though, and people are working to improve data quality and interoperability. Deep learning models are achieving human-level performance across a number of biomedical domains. In other clinical settings, it is expected that DL methods will augment clinicians and researchers.
The full potential of DL in healthcare has not been explored yet. If future deep learning algorithms can enable scientists to ask questions that they did not know how to ask, we can predict that biology and medicine will be transformed.
Photo by Tobias Fischer