1. Introduction
In Module 1, Professor Regina Barzilay explored the types of AI currently used in health care and their possible applications. Despite AI’s potential, however, the number of successful applications in health care remains low.
In this casebook, Dr. Constance Lehman reviews one successful use of AI; specifically, in the imaging sciences. Through this example, Dr. Lehman identifies some of the challenges currently faced by health care practitioners, and how AI can be used to address those challenges.
While AI does have potential to address certain challenges, you might still face various obstacles when attempting to implement AI in a clinical setting. These obstacles, and how AI might change the future of health care if they are overcome, are also discussed.
Burdens on health care professionals
- Overburdened health care professionals: There are a number of repetitive tasks that take up a significant amount of time for health care professionals.
- Inadequate and inconsistent medical imaging analysis: Human performance in analyzing cancer screenings can be inconsistent and limit the possible care that patients receive.
- Lack of access to medical imaging: Currently, global health care services suffer from a lack of access to radiological expertise.
Dr. Lehman explored how AI is able to do the repetitive tasks that would normally fall to health care professionals, allowing them to focus on tasks that benefit from human capabilities. This would reduce some of the burden on doctors, preventing burnout.
Beyond saving doctors’ time, there is the real potential to improve care, such as in applying AI to mammograms. While some specialized radiologists are very good at catching signs of cancer, general performance can be inconsistent. As such, the quality of care a patient receives is dependent on who treats them. AI has the potential to remove this inequality.
Furthermore, having an AI model interpret medical images means that the availability of specially trained radiologists would no longer be a prerequisite for patients to have accessible and reliable diagnostic imaging, as computers could read images globally.
Challenges in implementing AI in a health care setting
three potential challenges to implementing AI solutions:
- Overenthusiasm: Enthusiasm should be approached with caution when you aim to implement AI in health care. It is necessary to be methodical and realistic. Overemphasized and glorified claims, such as computers replacing all physicians, can create unrealistic predictions and a false sense of what AI will be able to do.
- Fear: The inverse of overenthusiasm, you should not let fearfulness unnecessarily block AI implementation. Overly restrictive guardrails may impede the potential benefits of AI in health care. For example, societal fear that computers will not care for patients in the same way a doctor can, may prevent AI’s potential from being realized at all.
- Managing change: Change management for AI in health care is a complex task. Implementing AI in health care should be managed carefully and involve all relevant stakeholders. Fearfulness and overenthusiasm can be addressed with careful change management. It is important to ensure that patients, doctors, and administrators are aware of how AI can be implemented safely and beneficially, moving forward with the correct guardrails while still delivering effective results.
A suggested method of navigating these problems is to start with simpler applications, targeting your use case on low-hanging fruit, so to speak. Subsequent applications can then be built up from these starting points. For example, Dr. Lehman and Professor Barzilay chose to tackle the problem of unnecessary breast surgeries. The aim was to develop an algorithm that could discern between women who needed surgery and women who could safely avoid surgery. The AI algorithm performed at a far higher level of accuracy than traditional methods used for determining which women should have surgery.
A similar approach was applied to determining breast density. Normally, a radiologist would make a subjective assessment as to the density of a woman’s breast tissue based on medical imaging. However, due to the variability among radiologists, this can result in widely different assessments. Once trained, the AI model can provide a consistent and accurate measure. As noted by Dr. Lehman, breast density assessment is already federally mandated, so it is important to be able to offer consistent assessments. Yet, as Professor Barzilay mentioned in Module 1, breast density itself is a poor biomarker and AI can be used to identify better ones.
Managing future changes in health care
Implementing AI will cause changes to health care and the nature of health care roles.
Dr. Lehman noted that AI technologies are tools: It is up to people to implement them, so that they address the challenges faced by patients and health care systems. It is also important to have a rigorous and evidence-based approach to mitigate both the fears and zeal surrounding AI, and to deliver better care.
Dr. Lehman noted that the implementation of AI will change the roles of health care professionals, but that these changes will allow for a more human-focused use of their time. If a radiologist, for example, is no longer required to manually process mammograms every day, they can focus on tasks that benefit most from their human skills and capabilities.
Technical perspectives on AI diagnostic and monitoring techniques
Having access to the full medical history of a patient’s risk factors would increase the accuracy of the diagnosis, but access to this information is not always assured, nor is it uniform across patients and institutions. In response, machine learning algorithms can be trained to fill the gaps through reasoned predictions about patients’ risk factors.
An algorithm that is able to fill in missing risk factors improves predictive power and enables functionality for a wide range of settings where not all risk factors have been collected. By adding in such customizations, you improve the predictive power of your model. These customizations can, moreover, be incorporated into off-the-shelf models. You can use a standard model, assess performance, and gauge what level of performance would be sufficient for a broader clinical context and what would require further customization.
Casebook: AI-enabled patient monitoring
Revolutionizing patient monitoring
Medical processes require that vast amounts of data are recorded about patients. A large proportion of this data is biometric in nature, such as a record of breathing functionality or heart beats per minute (BPM).
Using wearables or analyzing patient behavior through other more manual processes, such as keeping a personal medical journal, may be rendered obsolete if a wireless alternative can fulfill the same job with more accurate results.
Capabilities of invisible monitoring technology
Current approaches to collecting health data include wearables and self-reporting. One disadvantage of wearables is that wearing them (as with the sleep stage example) can create an unnatural environment and bias measurements. Self-reporting is similarly biased as it can be difficult for a person to give objective and accurate self-assessments. For example, requiring a patient to track the effects of medication from their subjective perspective cannot provide accurate results.
Using machine learning and wireless signals, patient behavior can be remotely tracked. For example, you can track a patient’s gait and movements without the need for wearables. This is significant, as gait speed is an important metric in Parkinson’s, Huntington’s, ataxia, multiple sclerosis, and various other conditions. It is also used as a predictor of aggravations in CHF (congestive heart failure) and COPD (chronic obstructive pulmonary disease). Similarly, patient sleep patterns can be monitored in their own home without the need for invasive wearables or attending a sleep lab. Conditions associated with particular sleep patterns, such as depression, can therefore also be monitored wirelessly.
Implementing invisibles in patient care
The previous section discussed the types of capabilities that Professor Katabi’s AI patient monitoring device, Emerald, offers: the monitoring of gait speed, sleep patterns, breathing, heart rate, and potentially more. Now that you have been introduced to wireless tracking’s theoretical foundation, Section 3 considers the results of implementing these capabilities in patient care.
Patients with Parkinson’s, Alzheimer’s, COPD, depression, or anxiety can have their behaviors monitored and based on these observations, certain inferences can be made. An erratic gait, for example, can be used to infer difficulties with balance and stability. With enough data collected, behavioral patterns can also be observed.
Figure 2 shows an Alzheimer’s patient’s activities over the period of two months (Kabalec et al. 2019). Each band of the circle denotes the timespan of one day, with each segment of the circle denoting a three-hour interval.
Conclusion
In this casebook, you have explored the capabilities and potential of wireless patient-monitoring technology. Such AI-powered approaches unlock a range of potential applications for health care. From improved accuracy and reduced intrusion while recording patient data to gauging the effectiveness of medication, invisibles will revolutionize approaches to monitoring patients in health care.
