Deep Learning and Artificial Intelligence in Healthcare

Artificial intelligence (AI) refers to automating intellectual tasks normally undertaken by humans. There are several well-established methods that enable machines to achieve AI, including traditional machine learning (ML) methods such as random forests and support vector machines. However, in recent years, a particular type of ML model, known as deep learning (DL), has gained significant traction and been able to achieve state-of-the-art results in many application domains including healthcare. Deep learning refers to the use of large neural networks to learn complex nonlinear patterns from a labeled data set.   

Healthcare is one of the most promising and influential application areas for deep learning. Some of the successful clinical use cases of DL include medical image analysis and natural language processing of medical records to improve disease diagnosis and prediction, identifying new biomarkers and enhancing drug discovery. Deep learning also has great potential for revolutionizing the operational aspects of healthcare delivery by augmenting healthcare operations efficiency, improving patient flow in hospitals and improving patient monitoring, among others. Researchers in the OR/MS community are in a unique position to take the lead in this emerging area and integrate the predictive power of DL with the prescriptive power of optimization methods to improve healthcare management.

There are several key advantages to DL, compared to traditional ML models, that will likely keep DL as the primary method for achieving AI in the years to come. First, unlike the traditional ML models that require careful feature engineering (often a manual process that requires extensive domain knowledge), deep learning is an end-to-end process, eliminating the need for feature engineering. Second, DL models benefit from economies of scale, whereby increasing the amount of training data adds little computational cost but often results in a notable performance improvement. Third, DL models are highly scalable because the mathematical operations needed to train these models can be parallelized on GPUs and TPUs. Next, DL models are well-suited for online learning becauuse they can be trained on additional data without the need to start from scratch. Finally, transfer learning enables DL models to be repurposed for a different task, essentially allowing the knowledge of a model to be reinvested and reused for a new problem.

In the healthcare domain, massive amounts of structured, unstructured and semistructured data are stored in hospital information systems (HIS) that include patient medical records as well as operational data such as patient flow, scheduling and resource utilization records. Fortunately, these data are often recorded and maintained in formats that are consistent across healthcare organizations as defined by a few major international standards. This allows for scalability and adaptivity of AI models for healthcare management – crucial to its long-lasting success.

Despite the theoretical success of AI models for various healthcare applications, lack of interpretability and regulatory challenges have been primary obstacles to the wide acceptance and implementation of deep learning models in everyday clinical practice. In a healthcare setting, clinicians and hospital managers often need to understand how a model is making predictions and assess its face validity and consistency with the existing well-established body of medical knowledge before they can trust that model and use it in their practice. Interpretability of a complicated model can also help identify its potential vulnerabilities, thereby improving its reliability. In recent years, explainable AI has been an active area of research, and several methods to infer explanations from complex models have been developed. These methods are often focused on identifying the attributes of input data that have the greatest influence on model predictions and employ feature analysis to shed light on the rationale of the model predictions.

With many AI models already matching or surpassing expert-level performance, the future of AI in healthcare is both exciting and alarming. It is exciting because leveraging such AI solutions can dramatically improve the quality of care and enhance patient safety by reducing human errors and decreasing provider fatigue from repeating routine clinical tasks. AI also has the potential to improve access to care by bringing much-needed clinical expertise to resource-limited and remote areas where access to specialists is sparse. Moreover, AI is likely to improve healthcare delivery by streamlining healthcare operations and reducing patient length of stay, opening additional healthcare capacity and allowing hospitals to accommodate more patients, thus improving the profitability of hospitals and providing economic incentives for making AI-related investments.

On the other hand, some healthcare tasks will inevitably be replaced with AI models, which subsequently can redefine the role of healthcare providers and impact healthcare operations. AI will allow healthcare providers to focus on tasks with the greatest positive impact. For example, radiologists may spend more time explaining radiology findings to the patients rather than reviewing radiology images. In the short-term, however, without clear regulation, clinicians are likely to act as learned intermediaries, evaluating and endorsing the decisions suggested by AI models.