If you have been on the Internet at all within the last 2 years...

You have probably noticed the impact of AI on the modern world. Its influence has seeped into everything, including business, writing, art, and more recently, the medical field.

Personally, I think that the phrase "AI", or artificial intelligence, is a misnomer. What "AI" actually is, in essence, is the intersection of incredibly advanced mathematical models with implementation as computational models in order to solve tasks that are tedious or costly to accomplish using humans. When people talk about AI, they are usually talking about language models (e.g. ChatGPT or Google Gemini). However, as I have said, these models are not intelligent at all. As in, these language models are not actually capable of what we call human reasoning. What language models actually do can be summarized as extremely sophisticated autocomplete using statistical prediction software that tries to predict what words are the most likely to come after a certain input of text. So these models are incredibly good at sounding correct, as they have been trained on heaps and heaps of textbooks, papers, articles, and other high-quality textual data. However, they can be convinced that two plus two equals five and that the sky is red. For this reason, I do not in good faith like using the phrase "AI" to refer to what can be more accurately referred to as machine learning or deep learning. What we call AI models is not "just a bunch of if statements" nor is it real intelligence, but mathematical models capable of taking a humanly perceptible data, processing it into matrices that can be computed on using calculus, and outputting data that is useful to humans.

AI in Healthcare

With that tangent aside, I would like to discuss how machine learning is used in the medical field, which is a topic I have become interested in since last year. In the medical field, the most typical use of ML is with classification algorithms and computer vision. Classification algorithms take an input of some sort of data and classify that data as some discrete output (e.g. has Alzheimer's vs. does not have Alzheimer's), which is very useful for diagnosis among other things. Computer vision is very broad as a field, but it includes image generation, image processing, image restoration and distinguishing certain objects from an image from other parts, which is called instance segmentation. Computer vision is useful for modelling and operating on various types of medical data, including CT scans, MRIs, and X-rays.

I believe that ML is currently an underutilized tool in the medical field, as it represents an opportunity to greatly democratize quality healthcare for underprivileged populations, including those in third-world countries who may not have access to highly trained professionals needed to diagnose certain conditions. ML can improve many lives, as ML can be implemented and used to model the progression of diseases for more cheaply than hiring doctors, and also can be more easily personalized to a person's individual health data.

Applications

I would say one of the most promising applications of machine learning in healthcare is in medical diagnostics. Traditionally, diagnosing conditions such as cancers or neurological disorders has historically required the expertise of highly trained specialists who can interpret complex medical images or a group of symptoms. However, ML algorithms can be trained to recognize patterns in medical data that may be too subtle for humans to notice. For instance, deep learning models have been developed to analyze mammograms and detect breast cancer with accuracy comparable to that of radiologists.

These advancements are not just about matching human performance; they are about scaling expertise. In regions where medical professionals are scarce, ML diagnostic tools can serve as a second opinion, ensuring that more patients receive accurate diagnoses and appropriate treatment. ML models can also continuously learn and improve from new data, potentially surpassing human diagnostic capabilities over time. I believe this fact is why medical ML has so much potential. As more and more high quality data is fed into models and more research is done, medical ML models will only get more accurate, more versatile, and cheaper to use in comparison to traditional methods.

Problems in the Field

Of course, with any emerging technology, there are always problems with its implementation and use. For ML in healthcare, there are significant problems which need to be solved. One of the most difficult problems as of right now is explainability. Essentially, explainability in the AI field is how easily researchers and programmers can understand how ML models reach their output. As it stands right now, not even the most advanced researchers in explainable AI have been able to figure out how to make the outputs of models explainable. Put simply, almost all advanced models are essentially a "black box" in which information is entered and you receive an output without any real knowledge of how the model reached its conclusion. This is extremely problematic, as there is no way to check the model's conclusion without running the tests manually, which defeats the purpose of using ML for these tasks.

There is also the problem that data for training models is often limited, as quality data for training can be very difficult to create without ethical issues. In the research lab that I work in, much of the data that the lab requires to train models, such as CT scans, is usually either very expensive or missing qualities that is vital to our projects. For this reason, there has to be a better system of creating and using data in medical ML than what we have now.

I personally believe that the problems above are solvable, but not without great effort and research. However, there will have to be a significant amount of collaboration between the ML research community and the healthcare field. Also, regulatory bodies will need to establish clear ethical guidelines for using ML in healthcare in order to actually address data privacy, informed consent, and accountability. Otherwise, it is unlikely that the use of ML in healthcare will ever see widespread use, as using unethically trained models may ruin the integrity of the researchers who develop them.

One of my forecasts that I think will happen within the next 20 years is the widespread integration of healthcare with machine learning. I believe that machine learning presents too much of a useful tool to be ignored, especially when they can be implemented at scale. As long as they are properly trained on good data, ML models for healthcare could one day outperform even trained medical professionals in diagnostic accuracy. I also predict that data harvesting in the medical field for training ML models will also become a subfield of data science, as medical data in general is extremely valuable, not only for training ML models, but for research as well.