Article by Madhumitha Rabindranath

Graphic design by Michie (Xingyu) Wu

Artificial intelligence (AI), machine learning, and big data are predominantly associated with tech industries; however, these tools are becoming increasingly applicable to other disciplines, especially medicine. Institutions are offering workshops and courses about AI and the University of Toronto recently founded the Temerty Center for AI Research and Education in Medicine (T-CAIREM) to drive AI innovation in healthcare.¹ For individuals who are removed from the world of computer science, AI concepts can be overwhelming.

What is AI?

According to IBM, AI “leverages computers and machines to mimic the problem-solving and decision-making capabilities of the human mind”.² To put it simply, companies continually mine for large datasets that AI can use to parse and make meaningful decisions, solving more complex and interesting problems.

There are two forms of AI – narrow and general AI.² Most applications use narrow AI where models perform specific tasks with high accuracy. For example, virtual chatbots are algorithms used to resolve routine issues. Conversely, general AI works to mimic the human mind such as achieving self-awareness. Examples of these are found in science-fiction such as in the film, I, Robot but have not been achieved in the real world.

Within AI exists different subclasses that are implemented for specific functions. Machine learning (ML) is a set of algorithms that are trained to analyze large datasets and create predictive models. Deep learning (DL) has a similar function to ML but is useful for unstructured data such as imaging and requires less human input for training. Other types include natural language processing and robotics. With diverse functionalities, AI is a promising tool that has the potential to revolutionize our world.

Why is it a big deal now?

Although the buzz around AI is recent, its concept was first envisioned in the 1950s. John McCarthy and Alan Turing, two of the founding fathers of AI, coined the term at a Dartmouth summer conference, spearheading AI research.^3,4 While many scientists tried to build complex machines, there were limitations to their implementation such as cost and low computational power.^5,6 As technology improved, the gains in computational power and data storage were matched with lower costs, making AI models more attainable. 

What are some clinical applications of AI?

AI applications in medicine are endless, ranging from diagnostic and predictive models to assistive therapies. The general principle of building AI models requires training algorithms with large, labeled datasets and testing the model on a novel cohort, validating its performance against gold standards. The process is long as models require continuous fine-tuning; thus, most medical applications are still in development. The following are some clinical applications of AI.

Electronic Patient Records

Electronic patient records (EPRs) have large amounts of clinical information which can be useful for research and clinical decision-making. However, EPRs are rarely structured and are difficult to parse through. One way to tackle this issue is to use natural language processing algorithms that flag keywords during patient-physician conversations and automatically fill in relevant EPR sections.⁷ This would facilitate standardized record-keeping and better patient-clinician interactions, reducing the burden of maintaining records on clinicians. Intelligent record-keeping systems can also assist clinicians in flagging patients for review and potentially interact with wearable technology to enable constant patient management. 

Histopathology and Medical Imaging

To improve diagnostic workflows, DL can be used to learn specific features of medical imaging or pathology slides which can assist clinicians in providing diagnoses. Several studies have shown that using DL to diagnose various pathologies such as cancer type and fibrosis is comparable to clinicians.^8–10 Some tools are also undergoing approval for clinical use. For example, Medo.AI recently received FDA approval for their AI platform that detects hip dysplasia, common in infants, using ultrasound images.¹¹ 

What are some issues with AI?

Despite its advantages, there are some challenges to implementing AI in clinical settings. Many clinicians are enthusiastic about the implications of using AI, but its integration is stalled by the limitations of translating these models to the clinic. The following are some issues of AI in healthcare and how they can be overcome.

The “black box” problem

The flexibility of using AI for diverse problems comes at a cost. Dubbed the “black box” problem, we are unable to explain AI decision-making as we can only make educated inferences, despite many AI models’ outputs matching clinical expertise. Without the means to understand AI decision-making, it is risky to use these tools in the clinic especially concerning patients’ lives. Fortunately, efforts to create more explainable AI models are underway but there are concerns about affecting AI performance.¹² Another way to tackle this issue is to have domain-specific guidelines for interpretability.¹³ For example, in computer vision (a subclass of DL for image analysis), models can learn specific prototypes for parts of the image that are important for classification, and new images testing the model are classified based on learned prototypes.¹³ These prototypes can provide more interpretability of its learning process, increasing our confidence in AI decision-making. 

Data Security and Safety

Another concern stems from data security issues and biases in AI models. For generalizable and successful modeling, AI tools require large amounts of data, requiring researchers to use open-source datasets. However, in clinical applications, hospitals will need to make EPRs widely available and share data across institutions, posing serious data security concerns. Many private companies are also interested in healthcare applications of AI and can receive access to EPRs, infringing on patients’ privacy rights.¹⁴ Hence, anonymized data may be useful in these situations; however, reidentification is possible using these algorithms.¹⁴ To ensure ethical compliance, regulatory bodies can provide oversight. For example, the FDA is recently providing a new framework to regulate clinical applications of AI models.¹²

Final Verdict?

Despite these challenges, AI is a promising tool that has the capacity to transform medicine, both in research and clinic. Extensive validation studies are required before these algorithms are deployed to the clinic. With increasing innovation in this field, AI will help tackle complex clinical problems and provide decisions that consider this complexity. AI, ML, and its variants will truly bring medicine to the digital era and provide opportunities to revolutionize healthcare.

References

1. U of T launches new Temerty Centre for AI Research and Education in Medicine [Internet]. [cited 2021 Sep 27]. Available from: https://tcairem.utoronto.ca/news/u-t-launches-new-temerty-centre-ai-research-and-education-medicine

2. What is Artificial Intelligence (AI)? [Internet]. 2021 [cited 2021 Sep 28]. Available from: https://www.ibm.com/cloud/learn/what-is-artificial-intelligence

3. Amisha, Malik P, Pathania M, et al. Overview of artificial intelligence in medicine. J Fam Med Prim Care. 2019;8:2328–31. 

4. Artificial Intelligence (AI) Coined at Dartmouth [Internet]. Celebrate Our 250th. 2018 [cited 2021 Sep 28]. Available from: https://250.dartmouth.edu/highlights/artificial-intelligence-ai-coined-dartmouth

5. The History of Artificial Intelligence [Internet]. Science in the News. 2017 [cited 2021 Sep 28]. Available from: https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/

6. Early Popular Computers, 1950 – 1970 – Engineering and Technology History Wiki [Internet]. [cited 2021 Sep 28]. Available from: http://ethw.org/Early_Popular_Computers,_1950_-_1970

7. Willyard C. Can AI Fix Medical Records? Nature. 2019;576:S59–62. 

8. Ghahremani P, Li Y, Kaufman A, et al. DeepLIIF: Deep Learning-Inferred Multiplex ImmunoFluorescence for IHC Quantification [Internet]. 2021 May [cited 2021 Sep 24] p. 2021.05.01.442219. Available from: https://www.biorxiv.org/content/10.1101/2021.05.01.442219v1

9. Yu Y, Wang J, Ng CW, et al. Deep learning enables automated scoring of liver fibrosis stages. Sci Rep. 2018;8:16016. 

10. Dermatologist-level classification of skin cancer with deep neural networks | Nature [Internet]. [cited 2021 Jan 23]. Available from: https://www.nature.com/articles/nature21056

11. MEDO.AI. MEDO.ai receives FDA approval to automatically detect hip dysplasia, preventing the leading cause of early hip osteoarthritis and hip replacement surgery. GlobeNewswire News Room. 2020. 

12. Kelly CJ, Karthikesalingam A, Suleyman M, et al. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17:195. 

13. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1:206–15. 

14. Murdoch B. Privacy and artificial intelligence: challenges for protecting health information in a new era. BMC Med Ethics. 2021;22:122.