Advances in Artificial Intelligence

Artificial Intelligence (AI) is being used in almost every aspect of science and engineering. In AI tasks, computers perform tasks that have normally required human intelligence. AI has shown significant progress in speech recognition and natural language processing

In 2017, JASON published a study that investigated how AI will shape the future of public health, community health, and healthcare delivery . One finding from the study was that AI can perform clinical diagnostics on medical images equal to expert clinician diagnosis, on specific cases.

Although there is an abundance of healthcare data available, it is often not easily available to create algorithms that aid in data-driven decisions. Health data is privacy protected therefore it makes sharing the data very difficult compared to other datasets. The lack of interoperability between electronic health records also poses an issue for the collection of data .

Types of AI

For as long as computers have been in existence, some form of AI has been in existence as well, given a computer's ability to mimic human intelligence. However, as the technology has grown, the term AI has come to encompass several different types of technology, which is important to distinguish:

Symbolic AI

Symbolic AI is an effort to explicitly codify human knowledge into some form of computer representation. These often existed in the form of conditional logic (i.e. if-then statements) and decision trees, and for many decades, this was essentially the only form of artificial intelligence available. Although readily explainable in the sense that any user could follow the built-in logic and understand why a model may have reached a particular result, it was inflexible, and often could only be utilized in very specific use cases.

Deep Learning / Machine Learning

By the 2010's, there was a significant growth both in the size of datasets along with the processing power of computers. Researchers realized that algorithms could be used to process and analyze these datasets in order discern meaningful knowledge. Commonly used algorithms include regression models, support vector machines (SVM), K-nearest neighbor (KNN), decision trees (similar to those used in symbolic AI but much more complicated), and neural networks. Broadly speaking, strategies include:

Supervised Learning

The model is provided explicitly labeled output training data such that it can learn to predict this output in new, unseen data.

Unsupervised Learning

The model is not provided explicitly labeled output training data, and instead identifies patterns and relationships in the data.

Generative AI

Another advance in artificial intelligence (AI) is generative AI. The key word is generative, which intimates that something novel is being generated. Generative AI technology relies on deep-learning algorithms, oftentimes neural networks, to generate new content (text, audio, etc) from a wide variety of sources. These sources can be both unstructured (clinical notes, recordings, images (and structured data sets (numerical values).

Some of the excitement behind this stems from the ability to create an aggregated summary from all the disparate sources of information present in a person’s medical record, in a short time frame. This could have the potential to save providers lots of time and effort procuring the right information, which they normally must do every time they see a patient. This allows for providers to focus most of their time on focus on the patient and tend to their care plans. Consequently, this can have an impact on reducing burnout as well, given that many clinicians indicate time sifting through the electronic medical record as a key factor in their fatigue. Another benefit is that generative AI allows for the transition to person-centered care. Many patients complain that clinicians spend more time behind a computer than working with them during encounters; generative AI champions believe this technology can flip the narrative. While there is a lot of positivity around this topic, generative AI’s output needs to be vetted for accuracy. Some type of proof reading is needed to ensure the final product is accurate and meets the standards of clinical documentation, for example.

Some industry focused applications of generative AI are making headwinds. For payors, member engagement and personalized outreach could be streamlined on the back end by reducing the administrative burden backlog. Additionally, prior authorization claims could be addressed in a speedier fashion. For health systems, automation of discharge summaries and streamlining administrative burdens there as well can have huge impact on care coordination, quality, and safety.

Medical Imaging / Diagnostics

The JASON study cited a study where automated retinal image analysis was conducted. The algorithm was trained using a dataset of over 100,000 images. The dataset was reviewed by 3-7 ophthalmologists. The results from the algorithm were comparable to results derived from manual assessment . This study demonstrates the potential of increased AI driven outcomes if accessibility is granted to valuable datasets. If this work continues the technology could be used to support clinical decision making, reduce the costs of a manual ophthalmologist assessment, or extend medical services to underserved populations.

Another study that was performed involved a training set of over 125,000 dermatologists labeled images from 18 online repositories. The study used a convolution neural network algorithm to classify melanoma. The algorithm performed similarly to a dermatologist diagnosis .

Both studies demonstrate that an algorithm can perform levels similar to its training set.

Patient Data

There are currently may smartphone attachments and apps for monitoring one’s personal health. These devices empower individuals to monitor and understand their health; create large amounts of data that could be used for AI applications; and capture health data that can be shared with clinicians and researchers. If patients were to provide access to mobile health data, this could enhance the research community’s ability to build more insights into public health through AI.

To collect the data, participants must provide consent. One study was able to do this through the creation of an app that was available through the Apple Store. The app was created for individuals with Parkinson disease. The app explained the study, those who were eligible to participate were asked if they wanted to share their data. Of the 12,200 eligible participants, 78% agreed to share their data . This is one way to collect patient data in order to further AI techniques to study health data, but patients must consent and there should be a clear plan as to how the data will be used.

Ambient AI

Another advance in AI is called ambient AI. This differs from generative AI in that ambient AI focuses on passively documenting and capturing real time information for patients. Examples of this include ambient AI with regards to note-taking during a patient encounter in the ambulatory setting. Another example includes information being derived in real-time from wearables and remote patient monitoring (RPM) devices. The important aspect of this is that the technology focuses on capturing the essence and purpose of a patient interaction in real-time. With regards to wearables and RPM, ambient AI provides us with more data points on the chronic conditions of patients, for example, which allows clinicians to have greater insight and create more nuanced care plans. Within the realm of clinical documentation, it will significantly reduce the documentation burden on providers without disrupting their normal workflows and allow them to be more in tune with their patient care responsibilities. This technology has advanced significantly over the last few years where now the passive documentation and capture of conversation can happen in real time.

Challenges of AI

There are a number of challenges that exist with AI models, particularly those involving machine learning and generative technologies.

Data Quality

Models are only as good as the data that are being used to train them. Availability of non-biased, high-quality data is challenging to come by, particularly when appropriate privacy and security issues are taken into account.

Modeling Challenges

The training of the model is an interplay between the algorithm and the training data; overfitting can occur when the model is too specific for the training data and is not broadly generalizable. Many models are based on algorithms that are inherently inscrutable in nature, i.e. "black boxes", in that they cannot explain why they may have reached a particular conclusion. As noted above, bias often exists in the training data, and this bias is inevitably reflected in the model itself.

Implementation

There remains a dearth of evidence based on randomized clinical trials demonstrating effectiveness of AI applications in healthcare. Organizations will need to devote significant resources to implementation and effectiveness trials. Stakeholders, particularly front-line clinicians, need to be involved in every step of the process of AI integration. Once implemented, there also needs to be ongoing monitoring of model performance, including for "drift," which is when the patient population on which the AI model is being used begins to differ from the training data over time. Implementation also needs to consider thresholds around model use, as excessive false positives can lead to alert fatigue.

Ethical

There are many gray areas that will take time to clarify. Presently, it's unclear who is to be held accountable if an AI tool causes patient harm: is it the AI developer, the health system that implemented it, or the clinician who accepted its recommendation? Many organizations are only now beginning to develop appropriate governance structures and policies about appropriate AI use. Furthermore, data collection from patients need to revisit formal consent around use of their information for the purposes of training AI models.

Social

There are significant social implications to the use of AI. Will patients come to view medical expertise differently if they feel they are able to obtain it from a chatbot instead of a clinician? What is the role of a clinician moving forward as more and more AI is implemented? What is the workforce impact on the healthcare industry if AI can start replacing aspects of roles? These remain unanswered. Furthermore, healthcare systems will also need to think about equity and access to these new technologies across patient and clinician populations.

With proper awareness, thoughtfulness, and implementation, however, there remains tremendous potential for the use of AI in healthcare.

References

Elizabeth Gamino

Submitted by Suvrat Chandra

Submitted by Allen Chang