In medicine, the ability to predict the course of a disease provides physicians with crucial information that may impact the type and length of therapy prescribed. Unfortunately, many diseases do not follow a straightforward path of progression; it is one thing to say that the mean outcome of a disease is such and such, but it is another thing altogether to say whether this will play out in a patient exactly as expected. 

How do we get better at predicting clinical outcomes? The problem does not lie in a lack of data; data collection in many developed countries is highly advanced. Perhaps the real problem lies in our inherent limitation in processing enormous sums of data in a way that can eventually yield clinically meaningful results. 

However, there is a scientific development in the field of technology that can potentially solve this problem—artificial intelligence, otherwise known as machine learning. In a nutshell, the term “machine learning” is coined as such because it is a system that can learn from previous input and hence interpret future data more intelligently. 


Continue Reading

The beauty of artificial intelligence is that it is, in a sense, blind. It shows no favoritism towards a particular trend and demonstrates no bias towards a particular product. It takes in information as it is, analyzes it, and tells us what it means. 

In the journal Med, Adlung and colleagues wrote, “Machine learning is increasingly integrated into clinical practice, with applications ranging from pre-clinical data processing, bedside diagnosis assistance, patient stratification, treatment decision making, and early warning as part of primary and secondary prevention.” 

In order to determine the efficacy of machine learning in a clinical setting, it must first be tested against our current gold-standard capacities in terms of clinical suspicion, diagnosis, and epidemiological studies. In this regard, studies have demonstrated that machine learning performs at least as well as clinicians. 

Garnering Data From Endoscopic Images

One recent study looked at how artificial intelligence is used to predict the malignant potential of gastric gastrointestinal stromal tumors (GISTs).

One of the key tools used for the diagnosis and evaluation of GISTs is endoscopic ultrasonography. In most cases, the specimens obtained are inadequate for the assessment of mitotic activity, meaning they are limited in their ability to stratify malignant potential. But in the event that sufficient specimens are obtained, scientists have proposed a number of parameters that can be used to predict malignant potential, such as tumor size, presence of ulceration, irregular margins, hypoechoic halo, distorted shape, extraluminal growth patterns, and heterogeneity. 

Read more about GIST etiology 

Given the potential of machine learning, Seven and colleagues decided to investigate whether artificial intelligence could use previously obtained endoscopic ultrasonography images from cases of surgically resected gastric GISTs to predict malignant potential.

The research team looked into patients who underwent an endoscopic ultrasonography at a single tertiary referral center for the assessment of subepithelial lesions. They enrolled participants (N=55) who had a confirmed diagnosis of gastric GIST according to the Armed Forces Institute of Pathology (AFIP) criteria, underwent an endoscopic ultrasonography prior to or after surgery, and had endoscopic ultrasonography images stored in a digital format.

Pathology studies were conducted via conventional methodology for each sample to analyze their malignant potential. A total of 685 images were collected and analyzed. The artificial intelligence model used in this study was convolutional neural networks (CNNs), which is a deep learning model mostly used for image processing and computer vision. 

Read more about GIST patient education

In terms of the accuracy of artificial intelligence in predicting AFIP criteria, the overall sensitivity was 75%, specificity was 73%, positive predictive value was 68%, and the negative predictive value was 79%. In addition, the authors of the study wrote, “When the patients were divided into low-risk and high-risk groups, sensitivity increased to 99%, specificity to 99%, positive predictive value to 89%, negative predictive value to 89%, and accuracy to 99%.” 

Seven and colleagues concluded that the use of CNNs as a deep learning algorithm using endoscopic ultrasonography images was a highly useful tool in predicting the malignant potential of GISTs. In addition, they discovered that artificial intelligence performed well in terms of predicting a high mitotic index.

Incorporating AI Into Clinical Practice

This study is the first to investigate the utility of artificial intelligence in predicting the malignant potential and the mitotic index of GISTs. It hence serves as a milestone study to help us understand how artificial intelligence can be further incorporated into clinical research. 

Artificial intelligence is vast and complex; finding the right use for it in clinical medicine will inevitably require more research. In addition, even if artificial intelligence meets certain clinical standards, it may still face regulatory hurdles for its implementation in clinical practice. 

However, the potential of artificial intelligence in medicine is undeniable.

“The extensive and unique contributions of [machine learning] systems to clinical decision making may generate a substantial impact but necessitate continued rigorous research and tackling of challenges and biases . . . transforming the data-driven evolution of precision medicine,” Adlung and colleagues wrote.

References

Adlung L, Cohen Y, Mor U, Elinav E. Machine learning in clinical decision makingMed (N Y). 2021;2(6):642-665. doi:10.1016/j.medj.2021.04.006

Seven G, Silahtaroglu G, Kochan K, Ince AT, Arici DS, Senturk H. Use of artificial intelligence in the prediction of malignant potential of gastric gastrointestinal stromal tumorsDig Dis Sci. 2022;67(1):273-281. doi:10.1007/s10620-021-06830-9