Last year, I was doing some data exploration on MIMIC IV, the largest open-source healthcare database produced by Beth Israel Hospital –one of Harvard’s teaching hospitals.
During my exploration, I wanted to see which diseases had the highest incidence over the years. I used something called the ICD codes –a universal/unique identifier for each disease– to organize a table that displays diseases from highest to lowest frequency. Below I attached just the first 5 rows of the table I made (i.e., the 5 most frequent diseases).
icd_code | icd_version | long_title | frequency |
---|---|---|---|
3893 | 9 | Venous catheterization, not elsewhere classified | 13928 |
02HV33Z | 10 | Insertion of Infusion Device into Superior Vena Cava, Percutaneous Approach | 10061 |
8938 | 9 | Other nonoperative respiratory measurements | 10041 |
3897 | 9 | Central venous catheter placement with guidance | 9843 |
8856 | 9 | Coronary arteriography using two catheters | 9043 |
966 | 9 | Enteral infusion of concentrated nutritional substances | 7811 |
As you can see, row 1, 2, 4, and 5 are all related to the Cardiovascular system, or more simply, the heart and vasculature. Not only are cardiovascular pathologies the most common, but they often affect some of the most sick patients in the hospital (e.g., chronic heart failure).
Right now, AI is a hot topic in cardiology. It makes sense because, if done right, AI tools in the field cardiology could address some of the highest burden and most devastating diseases in the hospital. Now that I’ve finished motivating the importance of this topic, I am to going to give a brief summary of exciting use cases showing up in the literature.
The buzz around AI applications in cardiology are mainly focused on using deep learning to provide personalized medical care for patients. The way this happens is by using large swathes of data from a patients profile on an EMR like EPIC and running deep learning models to predict outcomes like how likely is a person to get an infection after a procedure. The deep learning models allow for granular data points unique to each patient to be used as input for a tailored treatment plan.
The biggest roadblocks are legal in nature having to do with privacy laws, ethical concerns, bias, and the life or death implications of a mistake caused by AI tools. All these concerns are justifiable and need to be addressed, but it also must be acknowledged that the potential for AI to improve quality of care and save time and money, is severely hindered by these roadblocks. If you speak with anyone who works in a hospital setting, they will almost unanimously express some concern with the electronic medical record system. If you compare that software to the kind that is being leveraged on social media for marketing purposes, the disparity is night and day.
I am not arguing that the appropriate safeguards shouldn’t be implemented before widespread use of AI tools in cardiology, but it is becoming clear that people are skewing public opinion to have a disproportionately negative view of AI by overemphasizing errors that happen in a small minority of cases. A more honest assessment would be for people to compare the error rates of AI models to human error because this is ultimately the tradeoff.
Some exciting applications that are already being implemented in clinics and hospitals include the following:
- Using AI to predict cardiovascular outcomes of patients –related to personalized medicine we mentioned
- Using AI predictions using data from wearable devices to detect arrhythmias
- Prediction of outcomes for heart failure patients
- Accurate diagnosis of Coronary Artery Disease
- Improvements to at home screening methods
In addition to overcoming the legal issues and red taping involved in implementing AI medical technology, there are also several more technical problems. The first that I’ve encountered first-hand in any type of AI research that I’ve done is the problem of getting high quality data. If you’ve ever trained a ML model, you know that the best way to improve a model is by using good data. In fact, predictions are only as good as your data. This follows the famous dictum, ‘garbage in, garbage out’.
The second technical problem is generalizability. In order to generalize an AI model, it must be able to make accurate predictions on a dataset that it was not originally trained on. However, it is difficult to allow for this kind of testing to even occur due to strict data-sharing policies across institutions and a lack of a universal standard for which features will be collected in a dataset. Despite these challenges, great strides are underway in these areas, and I am very optimistic that AI tools in cardiology will significantly improve quality of care while saving patients time and money.
References
One response to “AI in Cardiology”
Very insightful article! When EMR was initialy introduced, there were tremendous amount of resistence, mostly from the medical professionals, such as myself, who did not fully understand computers and technology. Ultimately AI will be adopted in medicine just as it is being adopted into every other aspects of life. A major task is to convince busy physicians to spent extra time and effort to implement it in a meaningful and practical ways into patient care. We need more medical professionals who understand AI and computers.