Let me just share my slides, and then, oops, sorry, okay, all right, are you guys able to see my slides okay? Yes. Okay, perfect. So today we're going to be talking about artificial intelligence in GYN oncology. So just outline of this mini talk, we're going to go over some background on the basics of machine learning. We'll talk about examples of machine learning in GYN oncology using the electronic health records, cervical cancer, endometrial cancer, ovarian cancer, and then finish off with some ethical concerns, limitations, and future directions. As Dr. Bailey mentioned, I'm by no means any expert in artificial intelligence, but have a lot of research interest in it, and we'll try to give a high-level overview and share what I know. So as you all may know, artificial intelligence has been advancing rapidly and is now applied in military defense, transportation, finance, and increasingly in healthcare. The use of AI in oncology has received increasing attention over the last 20 years, and so it's important to understand the foundation of AI and how it functions, because we will be seeing more of it. So first, what is artificial intelligence? It is a machine that behaves in ways that would be called intelligent as if a human was behaving so. So for example, a robot that is able to respond to patients' questions about a patient's diagnosis. Machine learning is a program that has the ability to learn and improve from experience without the creator having to manually alter the program to improve. And then natural language processing is a subset of machine learning that is specifically designed to understand spoken, written, or typed text. For example, when physicians use a microphone to record what they are saying, then translate it to typed text in their documentation. To make predictions when applied to new problems, machine learning and natural language processing use two principles, unsupervised learning and supervised learning. Very briefly, unsupervised learning is when you solve a problem and the algorithm doesn't have the outcome of interest. So for example, patients are grouped together based on common characteristics that describe them. The Cancer Genome Atlas used this for genetic alterations in endometrial cancer, and that led to identification of poly-P53 mutations, grouping by common genetic changes rather than their cancer histology and leading to better accuracy and prediction of survival. Supervised learning is where you have, when solving a problem, the algorithm knows the outcome of interest. So the algorithm tries to find the combination of characteristics that most accurately predicts the outcome of interest. This includes neural networks, random forest plots, and supervised learning is the most common type of machine learning used in healthcare. So now turning to examples, first we look at the electronic health record. So in the IBM Watson Health Study, machine learning reduced the time to screen patients for clinical trials by 80% by extracting electronic health record information and establishing clinical trial eligibility. So in the future, it's feasible that machine learning would be able to match clinical trials to eligible patients according to demographic, clinical pathologic, and molecular characteristics at a system-wide or national level. There was a study at UCSF that found that a machine learning algorithm incorporating patient charts was able to effectively predict readmission in hospital mortality and hospital length of stay. So if you knew that a patient was at a higher risk of readmission, for example, you would be able to offer nurse visits at home, more detailed instructions on discharge, and so forth, tailoring your care to that specific patient. Turning our attention to cervical cancer, Hewitt et al. used neural networks to determine the presence of severe dysplasia in colposcopic images. And that method actually outperformed traditional cervicography. So it is reasonable to believe that the application of machine learning could improve the see and treat approach in low-resource settings without pathology expertise or reliable continuity of care where the algorithm tells you it's high-grade dysplasia and you proceed with the leap, for example. Neural networks have been applied to predict survival after radical hysterectomy and to identify those at highest risk of death after recurrence. Because early discussion for palliative care is beneficial to reduce aggressive care, objective assessment of survival time could add more concrete information in patient counseling. Others have used different machine learning algorithms to predict complications after radical hysterectomy and decrease radiation toxicity. And then machine learning has also been used to improve the detection of metastatic lymph nodes on MRI. In the future, machine learning may be used to help you select who would benefit most from adjuvant chemotherapy, radiation, or immunotherapy. Then looking at endometrial cancer, one of the key applications of machine learning has been to differentiate benign hyperplastic and malignant endometrial lesions using neural networks. In surgery, they've even used machine learning to diagnose endometrial cancer based on hysteroscopic images alone. And continuing to develop these sort of algorithms may add additional assistance in pathologic diagnosis just because, as we know, classification of high-grade tumors can sometimes be unreliable or not reproducible. Another significant aspect of endometrial cancer management is accurate surgical staging. But as we know, many endometrial cancer patients have substantial comorbidities including obesity, which can preclude safe surgical staging. Using a combination of histology and features on pathology, machine learning algorithms have been able to predict metastatic disease in pelvic lymph nodes with nearly 85% accuracy. And then when you add in MRI imaging as well, then those algorithms can predict the presence of metastatic disease even in normal appearing lymph nodes. So this is something you select who to consider maybe performing surgical staging on or having that additional information even when you were unable to do the full complete surgical staging. Artificial intelligence is applicable in guiding treatment decisions as well for endometrial cancer. So applying random survival force to a series of patients with uterine serous cancers, it predicted that almost 30% of stage 1A patients may not need adjuvant treatment by grouping them into low, medium, and high risk groups. Another machine learning generated molecular score was shown to predict systemic treatment responses in patients with advanced uterine serous cancers. And then finally, for ovarian cancer, machine learning has been used in diagnosis, surgery, treatment, and prognosis. A neural network using menopausal status, CA-125, and various features on imaging was able to predict malignancy. In 2016, there was a new neural network algorithm that used some novel inputs, including image texture and pixel arrangement within the mass, which demonstrated a 99% accuracy for diagnosing malignant adnexal masses on ultrasound. And then in ovarian cancer surgery, machine learning algorithms have been evaluated for their ability to forecast optimal cytoreduction. So there was one score by Kawakami et al. that used serum proteins at diagnosis to predict residual disease. And then another group used a combination of preoperative CA-125, age, suspected stage, histologic grade, and type. And both algorithms predicted cytoreductive outcome accurately over 60% of the time. And then if you incorporate radiologic imaging within these machine learning algorithms, potentially you would be able to further optimize patient selection for radical surgery versus neoadjuvant chemotherapy. Other studies have looked at predicting response to frontline chemotherapy based on clinical information and protein markers. There was a group that looked at serum biomarker combination with clinical features to detect residual cancer at remission and estimate time to recurrence. They've also looked at various predictors of patient prognosis and platinum resistance in ovarian cancer as well. And so it's really exciting how these machine learning algorithms are able to use data from clinical information, radiologic, genomic findings to provide unique, noninvasive, real-time personalized treatment approach for patients. So taking a step back, there are multiple ethical concerns regarding the application of machine learning to healthcare. So one of the major criticisms is that machine learning is a black box. As data goes in and interpretation comes out, but the mechanism of how the program derived the solution is unclear. There are also a lot of legal and policy issues that arise with respect to product liability, malpractice, who's responsible when there's an error that is caused by a machine learning algorithm. And then there's the issue of selection biases and the data sets used to develop machine learning algorithms. So that can result in propagating the biases of the data set when applied to underrepresented patient populations. And then there are other ethical issues of data ownership, security of the data and patient privacy. And then finally, the quality of the outcome depends on the quality of the available information. When machine learning algorithms are compared to randomized controlled settings, they often underperform. And because large data sets vary in quality, it's just important to remember garbage in and garbage out. And so it is of utmost importance to design high quality clinically validated and ethical machine learning algorithms, but there is a need for high quality data, standardized interpretation and validation procedures used in a randomized trial. And then on a more national level, or at least within the US, FDA approval and regulation of the algorithms before we use them clinically. But overall, the future of AI and Gynon shows great promise. Here's just an example pathway for a cancer patient. And each of these touch points at prevention, screening, diagnosis, and so forth shows a critical series of decisions for oncologists and patients to make and yields a use case for AI that could provide incremental benefit. As our biological knowledge base and data streams grow in clinical oncology, these machine learning algorithms could be used to provide more and more precise patient groups with patterns that guide their specific treatment for the next unseen patient. And so basically, as we get more data, optimal cancer care, the care that results in the best survival and quality of life for a patient will become precision care where we're able to really tailor it for that next patient in an even more specific kind of guided fashion. And so just to conclude here, so AI has been shown to enhance diagnosis, refine clinical decision-making and advance personalized therapies in gynecologic cancer. However, there is a need to overcome challenges related to data transparency, quality and interpretation, and that really requires incorporation of these key features of any model that we use in health care of usability, validity and utility into our model design. And that requires prospective trials that incorporate these novel AI driven algorithms for diagnosis and treatment guidance and will likely go on to change not only like health care, but gynecological oncology care as well. All right, and I'll just open it up if we have any time for questions. I know we're running a little bit.

artificial intelligence

gynecologic oncology

machine learning

cervical cancer screening

personalized therapies