Machine learning in heart valve surgery

La semplicità è l'ultima sofisticazione - Simplicity is the ultimate sophistication - Leonardo Da Vinci Machine learning (ML) is a machine's ability to improve its performance based on previous results. The founding principle of ML was developed through computing and fast digital processing [1]. Moore's Law is the nidus behind ML, stating that computing power doubles every 2 years, bringing efficient speed and hence reduction of cost [2]. There are 3 ways that Artificial intelligence (AI) can help heart valve surgery clinically in the future: automation of the process of the risk, gaining insightful new information from extensive data analysis such as predicting who benefits from mechanical versus biological valve while help guides personal decision for both patient and operator based on real-time data using non-linear methods. Currently, our heart surgical data are collected in a low dimensional manner, mostly binary with varied completeness of data. Given the manual input and handling of heart valve data, AI can accommodate transferring data from one media to the other while at the same time addressing failures and reading clinical reports through natural language processing [3]. The culmination of this is through cloud technology that allows for streamlining and merging datasets from all channels, including suppliers (hospital sites), customers (pharma, patients) and analysts. While allowing a concise and efficient infrastructure to be built, ML systems are somewhat limited in their ability to learn and evolve. The so-called AI-driven data analysis remains intriguing and holds the most potential as a wiser system. At this stage, it aids human decision-making in recognizing patterns, making intelligent inferences beyond natural language processing with generalization, cost function comparison of training data to the ground truth (i.e. data corresponding to the expected result of an algorithm). While all these processes complement our clinical training and decision-making, they have limitations, especially in any surgery. We explain why. Bodenhofer et al. [4] reported on ML-based risk profile classification in 2229 elective heart valve surgery patients. The study involved non-linear ML tools such as random forests (RF), which are many decision trees, and support vector machines with variations (SVM), which are a combination of supervised learning models and associated learning algorithms. Both RF and SVM help in the classification and regression analysis of data sets. The authors go about utilizing these tools and comparing them to the EuroSCORE predictor equations. In simple terms, SVM generates the distance to the classes' boundary while RF provides the probability of belonging to the class. Akin to statistical methods, the accuracy of ML is based on the quality and completeness of the data collected and the rigor of supervision in the variable annotation. The authors appropriately and convincingly utilized various accuracy methods, including classification accuracy, balanced accuracy and area under the receiver-operator characteristics curve. Bodenhofer et al. conclude that RF is more accurate than EuroSCORE. The term accuracy is defined as the quality or state of being correct or precise. However, EuroSCORE is precise and is recalibrated regularly given the imputation of data and time interval [5]. Each of us builds a personal predictor equation based on our cumulative experience, expertise and knowledge that I would coin as ‘Real Intelligence’ instead of AI. Nonetheless, the growing platform and framework of AI allow parameters and metrics to be generated based on non-linear relationships and convolutional neural networks within layers of information. Such depth of analysis for surgeons would complement our intrinsic personal severity assessment while being obliged to meet standard expected universal outcomes and national benchmarks. ML is great for research purposes and prediction before surgery because of its enhanced accuracy per this study in the EJCTS. Enhanced accuracy based on enhanced statistical significance may not, however, be clinically significant or incremental to the surgeon's prediction and intuition. Surgical data are surgeon and institution dependent, and while transferrable in concept, the ML developed here by Bodenhofer et al. needs further prospective validation in other centers. Risk profile in heart valve surgery outcomes depends on pathophysiological variables and surgical technique, and other unknown factors. These factors are part of ‘real intelligence’ and we as a society are always curious about finding these unknown factors. What may be lurking in large data sets to generate new information via ML is typically what experienced surgeons perhaps already know but did not digitalize. The lack of digitalization in our daily practices beyond the availability of resources is that we may not be aware of new onsets of information processing given our busy clinical commitments. However, given the timeline of ML and its future power, we should be aware of ML to improve our depth of data analysis and to find new associations among variables and its potential applicability in the operating room. Timing of data collected and their temporal relationship to other variables is currently absent in surgery. However, AI has the full potential to entertain fuzziness (non-linearity) with multi-layer data, including integration with imaging and other biophysical profiles. Nonetheless, the content of existing surgical data sets is quantitatively and quantitatively different to speech or visual object recognition, areas in which AI has thrived. ML used in EKG imaging analysis for example can be used to predict mortality [6]. If images and other biophysical profiles are integrated into our existing surgical valve databases, AI and ML approaches can be far more valuable than current logistic regression methods in heart valve research. The authors of this paper are to be congratulated for integrating surgery outcome analysis and ML [4]. Recently, a meta-analysis of using ML in cardiac surgery arrived at a similar conclusion but warrants further research and application [7]. However, this dynamic relation between AI and surgical risk profile predictor outcomes requires time to extract the data from the records and input it. Working out these statistical correlations incurs cost and prone to human errors. Both have to be considered before adopting ML approaches into our clinical outcome research domain for broader applicability. Furthermore, we anticipate that AI algorithms will help speed real breakthroughs to improve human health via biophysical integration, to bring this in real-time throughout the pre-, intra- and postoperative phases [8]. Scaling down is as important as scaling up in data informatics. What is meant by this is that if only our electronic databases can capture all events and analyze instantaneously using AI and ML to reduce cost and human error, the inferences will be revolutionary to our clinical practices. When this happens, it will be a scale-up, comprehensive, ethical and adding value to patient experience as new real-time standard for our bright future.