Raising the Standard: The Evolving Role of Artificial Intelligence in the Future of Surgery

| December 1, 2020

by Laura Hansen, MD; David Pechman, MD; and Dominick Gadaleta, MD, FACS, FASMBS

Dr. Hansen is a Fellow of Minimally Invasive Surgery at Northwell Health. Dr. Pechman is a Bariatric and Minimally Invasive Surgeon at South Shore University Hospital, Hofstra/Northwell in Bay Shore, New York and Assistant Professor of Surgery at Zucker School of Medicine at Hofstra/Northwell in Hempstead, New York. Dr. Gadaleta is Chair, Department of Surgery, South Shore University Hospital; Director, Metabolic and Bariatric Surgery, North Shore and South Shore University Hospitals, Northwell Health, Manhasset, New York; Associate Professor of Surgery, Zucker School of Medicine at Hofstra/Northwell in Hempstead, New York.

FUNDING: No funding was provided for this article.

DISCLOSURES: The authors report no conflicts of interest relevant to the content of this article.

Bariatric Times. 2020;17(12):18–19

Artificial intelligence (AI) is an emerging technology that might soon provide all surgeons with the ability to improve their performance in real time. AI, the study of algorithms that enable machines to perform cognitive functions, such as pattern recognition, problem solving, and decision-making, is becoming more clinically relevant as computer power improves. AI is a diverse field of study with many applications, both in medicine and other fields. It comprises multiple categories of algorithms that can be combined in different ways to achieve different goals. Four main subfields of AI are machine learning, artificial neural networks, natural language processing, and computer vision.

Machine learning (ML), enables computers to learn and make predictions by recognizing patterns within datasets. This can be accomplished in three ways: supervised learning, unsupervised learning, and reinforcement learning. In supervised learning, algorithms are provided a labeled dataset.1 For example, pictures of normal computed tomography (CT) scans labeled “normal” and CTs with free air labeled “free air” are given to the algorithm, and ML identifies characteristics specific to the categories and can make predictions about unlabeled pictures. This style of learning is useful to predict a known result or outcome. In unsupervised learning, an unlabeled dataset is provided, and the algorithm will separate the data into categories with similar characteristics. This is more useful to search for unknown patterns within data. Reinforcement learning, on the other hand, occurs when a computer attempts to accomplish a task while learning from its own successes and mistakes. Reinforcement learning is useful for automated tuning of predictions or actions, such as controlling the insulin delivered from an artificial pancreas. Overall, ML is capable of identifying subtle patterns in large datasets by allowing for more complex and multivariate analysis than conventional statistics.

Artificial neural networks (ANN), a subfield of machine learning, allow for powerful risk prediction. ANN process signals within layers of simple computational units and connections between the signals are parameterized with weights that change as the network learns. ANN analysis of patient data has outperformed traditional risk prediction models in multiple scenarios (APACHE II for pancreatitis severity2 and in hospital mortality after open AAA repair3).

Natural language processing (NLP) is another form of AI that is already being used in clinical applications. NLP is the ability to understand human language. By incorporating semantics and syntax into analyses of written content, NLP can infer meaning and sentiment. NLP has been used to identify words and phrases within operative reports, which, when modified by phrases from postoperative progress notes, has been able to predict anastomotic leak with 100 percent sensitivity and 72 percent specificity.4 NLP is also utilized in automated postoperative patient assessment telephone calls, which have the ability to identify adverse events and direct patients to seek further care.

Perhaps the form of AI most applicable to surgery is computer vision. Computer vision is the understanding of images and videos. Computers can be programmed to achieve human-level recognition of objects and scenes. Real-time analysis of laparoscopic video yields 92.8 percent accuracy in automated identification of the steps of a sleeve gastrectomy, and is capable of noting missing or unexpected steps such as prolonged control of bleeding.5 Through computer vision, AI can be leveraged to process huge amounts of surgical data to identify or predict adverse events in real-time for intraoperative clinical decision support. It has the capacity to prevent serious adverse events, such as cutting the ureter during a colectomy or the common bile duct during a cholecystectomy, similar to lane departure warning systems currently used in many cars.

There are, however, limitations to the application of AI in medicine. The accuracy of the data used in algorithms is paramount to the success of the algorithm’s ability to predict outcomes. For example, while ML provides a powerful tool to uncover subtle patterns in data that might be missed by traditional methods, ML analyses are limited by the types and accuracy of available data. Additionally, the clinical context for interpreting the rich variables found in big data is often missing.

Despite these limitations, AI is already being applied across medicine to improve outcomes. Error rates in detecting cancer-positive lymph nodes have been decreased from 3.4 percent to 0.5 percent by utilizing AI in analyzing pathology slides.6 There are multiple commercially available AI (deep neural network) augmented computer-aided detection (CAD) algorithms being used for mammography screening. Some of these AI CAD algorithms have been shown to reach the same performance level of radiologists in assessing screening mammograms.7 When combining the assessment of an AI CAD algorithm and a radiologist, more cancers are identified than when combining two independent radiologist reads.

Research in AI is demonstrating the future potential for algorithms to augment clinical care in diverse fields. Deep neural networks have been used to classify skin lesions on par with board-certified dermatologists.8 The algorithm was trained with 129,450 labeled pictures of 2,032 different diseases (an example of supervised learning), then tested with biopsy proven binary classification cases (keratinocyte carcinoma vs. benign seborrheic keratoses; malignant melanoma vs. benign nevi). The deep neural network algorithm performed as well as 21 board-certified dermatologists in accurately diagnosing the lesions. This suggests a potential role for AI in point-of-care diagnosis of skin lesions and therefore more universal access to expert diagnostic care.

By combining AI algorithms with innovative hardware, research engineers at Johns Hopkins developed a smart tissue autonomous robot (STAR) capable of performing a small bowel anastomosis autonomously.9 The robot consisted of a bedside robot arm with an articulated laparoscopic suturing tool, three-dimensional (3D) surface reconstruction, and a 3D visual tracking system using near-infrared fluorescent imaging. STAR was able to perform ex-vivo linear suturing of a longitudinal cut along a length of intestine, ex-vivo end-to-end anastomosis, and in-vivo end-to-end anastomosis of porcine small intestine. They were able to demonstrate improved suture consistency and higher pressures to leak, without lumen size reduction compared to open, laparoscopic, or robotic-assisted bowel anastomoses. However, the time for STAR to complete the anastomoses was significantly slower.

Because a lack of quality data can limit the predictive capabilities of AI, surgeons should work to create and expand clinical data registries to ensure the diversity of our patient populations are included. Datasets that include operative video and electronic medical record data from surgeons around the world would enable AI to uncover practices and surgical techniques that are associated with improved outcomes. Surgeons have the clinical insight to guide scientists to answer the right questions that will advance the field of AI in surgery. Future applications of AI in medicine and surgery to improve clinical outcomes are only limited by our imaginations and our participation in the process.

References

  1. Hashimoto DA, Rosman G, Rus D, Meireles OR. Artificial intelligence in surgery: promises and perils. Ann Surg. 2018;268(1):70–76.
  2. Mofidi R, Duff MD, Madhavan KK, et al. Identification of severe acute pancreatitis using an artificial neural network. Surgery. 2007;141(1):59–66.
  3. Monsalve-Torra A, Ruiz-Fernandez D, Marin-Alonso O, et al. Using machine learning methods for predicting inhospital mortality in patients undergoing open repair of abdominal aortic aneurysm. J Biomed Inform. 2016;62:195–201.
  4. Soguero-Ruiz C, Hindberg K, Rojo-Alvarez JL, et al. Support vector feature selection for early detection of anastomosis leakage from bag-of-words in electronic health records. IEEE J Biomed Health Inform. 2016;20(5):1404–1415.
  5. Hashimoto DA, Rosman G, Witkowski ER, et al. Computer vision analysis of intraoperative video: automated recognition of operative steps in laparoscopic sleeve gastrectomy. Ann Surg. 2019;270(3):414–421.
  6. Wang D, Khosla A, Gargeya R, et al. Deep learning for identifying metastatic breast cancer. https://arxiv.org/abs/1606.05718. Accessed September 27, 2020.
  7. Salim M, Wåhlin E, Dembrower K, et al. External evaluation of 3 commercial artificial intelligence algorithms for independent assessment of screening mammograms. JAMA Oncol. 2020;6(10):1581–1588.
  8. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115–118.
  9. Shademan A, Decker RS, Opfermann JD, et al. Supervised autonomous robotic soft tissue surgery. Sci Transl Med. 2016;8(337):337ra64.

Tags: , , , ,

Category: Past Articles, Raising the Standard

Comments are closed.

«
»