From the mathematical model to the patient: The scientific and human aspects of artificial intelligence in gastrointestinal surgery

Abstract

Recent medical literature shows that the application of artificial intelligence (AI) models in gastrointestinal pathology is an exponentially growing field, with promising models that show very high performances. Regarding inflammatory bowel disease (IBD), recent reviews demonstrate promising diagnostic and prognostic AI models. However, studies are generally at high risk of bias (especially in AI models that are image-based). The creation of specific AI models that improve diagnostic performance and allow the establishment of a general prognostic forecast in IBD is of great interest, as it may allow the stratification of patients into subgroups and, in turn, allow the creation of different diagnostic and therapeutic protocols for these patients. Regarding surgical models, predictive models of postoperative complications have shown great potential in large-scale studies. In this work, the authors present the development of a predictive algorithm for early post-surgical complications in Crohn's disease based on a Random Forest model with exceptional predictive ability for complications within the cohort. The present work, based on logical and reasoned, clinical, and applicable aspects, lays a solid foundation for future prospective work to further develop post-surgical prognostic tools for IBD. The next step is to develop in a prospective and multicenter way, a collaborative path to optimize this line of research and make it applicable to our patients.

Key Words: Survivor bias; Data analysis; Machine learning; Ethics; Critical thinking; Postsurgical; Complications; Inflammatory bowel disease; Gastrointestinal surgery

Core Tip: Recent medical literature shows that the application of artificial intelligence models in gastrointestinal pathology is an exponentially growing field. In this work, the authors present the development of a predictive algorithm for early post-surgical complications in Crohn's disease based on a Random Forest model with exceptional predictive ability for complications within the cohort. The present work, based on logical and reasoned, clinical, and applicable aspects lays, a solid foundation for future prospective work to further develop post-surgical prognostic tools for inflammatory bowel disease.

INTRODUCTION

These are times of change. Technological development has undergone exponential growth in recent decades, which has had direct repercussions on practically all aspects of our lives. And although healthcare moves at a different pace, marked by rigorous safety and ethical protocols, it has not escaped this change. The progressive computerization of our clinical records and the introduction of novel technology in the assistance of our patients, for example, are tangible realities. Although these changes in our clinical practice can undoubtedly be beneficial, they often constitute a challenge and require further research to optimize them[1].

As scientists, the basis of our work is data. A physician cannot work without data. And if there is a technological revolution these days that is pertinent to highlight, it is that of data. The computerization and digitalization that I mentioned previously, together with the appearance of multiple tools that make it possible to obtain more and more accurate data, mean that we are at a historic moment in which we can obtain the greatest amount of data, (as well as the most accurate data), from our patients. But data alone is of no value. Data acquire value at the expense of their correct interpretation. An illustrative example is the so-called survivor bias. This logical fallacy leads us to focus on people who pass a selection process while leaving out those who do not, which can lead to false conclusions. During World War II, for example, the United States government incurred this bias when analyzing aircraft returning from combat, recommending that damaged areas be reinforced. However, the interpretation was incorrect: It was the undamaged parts that should have been reinforced because the aircraft that suffered damage in those areas were those that did not return from combat. It was statistician Abraham Wald from the Statistical Research Group at Columbia University who analyzed the data in the right way and understood what was happening[2]. Therefore, Wald proposed that the Navy reinforce areas where the returning aircraft were unscathed. The data was there, it was just that it had not been interpreted correctly.

ARTIFICIAL INTELLIGENCE

And amid this information revolution, artificial intelligence (AI) emerges. But what is AI? According to Wikipedia, it is a discipline and a set of cognitive and intellectual capabilities expressed by computer systems or combinations of algorithms whose purpose is the creation of machines that mimic human intelligence to perform tasks, and that can improve as they gather information. And from this definition, I would like to highlight a few aspects: The first is that these computer capabilities mimic human intelligence. We can teach these tools to simulate our way of thinking, to “mimic” it, but never to replace it. The conception of original ideas in medical research, the "spark" inherent to our species that has led to unique and extraordinary discoveries is, in the opinion of the author of this text, inimitable. I think it is important that we encourage and safeguard critical thinking and the ability to produce original ideas, since one of the risks of the abuse of these models is precisely this, the loss of our "spark". The second is that these models improve as they accumulate and compile information. This is the basis of progress in science and medicine: Collaboration. I believe it is essential that strong international research groups are encouraged to develop these lines of research together to enhance their results and allow them to be extrapolated globally. Together we are more.

That said, I would like to turn my attention to the paper entitled "Predicting short-term major postoperative complications in intestinal resection for Crohn’s disease: A machine learning-based study”[3]. In this paper, the authors present the development of a predictive algorithm for early post-surgical complications in Crohn's disease based on a Random Forest model with exceptional predictive ability for complications within the cohort [area under the curve (AUC) = 0.965 in the training cohort, AUC = 0.924 in the validation cohort].

Concerning AI models and gastrointestinal pathology, recent medical literature shows that this is an exponentially growing field, which indirectly translates the existing interest in this area. We found promising models for different pathologies, such as gastric cancer, liver fibrosis and cirrhosis, gastrointestinal stromal tumors, and Barrett's esophagus, among others[4,5]. Overall, all these models (both diagnostic and prognostic) show very high performances that exceed the gold standards previously used. However, in most cases, they have been conceived for research purposes and their actual implementability in clinical practice is low or non-existent. Regarding AI models and inflammatory bowel disease (IBD), recent reviews demonstrate promising diagnostic and prognostic AI models. However, studies are generally at high risk of bias (especially in AI models that are image-based)[6]. The availability of specific AI models that improve diagnostic performance and allow the establishment of a general prognostic forecast in IBD is of great interest, as it may allow the stratification of patients into subgroups and in turn, allow the creation of different diagnostic and therapeutic protocols for these patients. This goes hand in hand with the general trend in our profession towards so-called personalized medicine.

Surgery has always been a complex area in scientific terms due to multiple factors. First, the variability between surgeons in terms of surgical technique: Although the current tendency is to protocolize and standardize surgical practice, many variables are difficult to quantify (from the pressure exerted on the sutures to the amount of monopolar energy used). Second, the difficulty in reflecting as objective data many of the intraoperative aspects: We can quantify the surgical time, but it is difficult to quantify the degree of peritoneal involvement in a patient with peritonitis or the degree of intestinal involvement in a diverticular disease. Nevertheless, despite the conceptual aspects previously explained, predictive models of postoperative complications have shown great potential in large-scale studies[7]. In this case, the authors have opted for the creation of a model with variables that are simple to collect and reproducible, which, in addition to making it rigorous, allows for possible external validation in the future. The Crohn's disease activity index (CDAI), the serum preoperative albumin, and the surgical duration are parameters that any surgeon can easily collect on their patients. On the other hand, it is pertinent to note that all the elements included have a strong biological plausibility in terms of post-surgical complications: A patient with hypoproteinaemia, high CDAI, or prolonged surgery is logically more susceptible to complications than a patient in a more favorable situation. In this case, except for urgent surgery, there are two modifiable factors such as CDAI and hypoalbuminemia, and a factor that the surgeon can keep in mind during the procedure such as surgical time.

Although there is a significant lack of knowledge regarding AI models, most of them are based on mathematical algorithms that, through iterations, create predictive models. This is the same principle as that of other statistical models we are already familiar with, such as logistic regression. In this article, the authors opted for a Random Forest type model. This is a reasonable and justified decision, given that previous work has demonstrated the superiority of these models to other traditional models such as logistic regression[8]. However, there are nuances to this, given that in certain cases it may be preferable to opt for a model such as logistic regression (for example, in cases where the noise variables are scarce). Random forest models have important advantages such as they can handle missing values and outliers, they can deal with classification and regression problems, and they can handle large amounts of data efficiently. However, it can be difficult and time-consuming to train them.

As I said at the beginning, we live in times of change, of rapid, almost vertiginous change. The amount of data we handle daily has become practically unmanageable, we are becoming dependent on technology. I think this is a mistake. I think we need to reflect long and hard about the direction we are heading in and realize that speed, while important, is not paramount. We must protect our critical thinking. We must reflect on the why of things. We must use technology as a resource within our reach, but not become dependent on it. We must look at the data and analyze it, and it is our responsibility to interpret it properly, no matter how many tools we have to assist us. When an anesthesiologist sees a desaturation on his patient's pulse oximetry monitor, the first thing he does is not to open the oxygen flow, but to look at the patient's finger to see if the sensor is correctly positioned: That is what defines us, the ability to think and reason. Feynman, a renowned physicist who laid the foundations of quantum mechanics, once said, "Direction is more important than speed." I couldn't agree more. The present work lays a solid foundation for future prospective work to further develop post-surgical prognostic tools for IBD. And it does so because it is based on logical and reasoned, clinical and applicable aspects. This must be the way forward for AI. The next step, as I mentioned, is to develop a collaborative path to optimize this line of research in a prospective and multicenter way and make it applicable to our patients. We must not forget that we do not work with machines, we work with patients.

CONCLUSION

AI-based predictive models for gastrointestinal surgical pathology, such as the commented work, show promising results. However, larger-scale prospective studies are needed for validation. These models cannot supplant human reasoning and the human mind, so they should be conceived as complementary tools in research and not as integral automated elements.

ACKNOWLEDGEMENTS

To Helen Williams, for her help in the linguistic revision of the manuscript.