AI has been around for decades, even in the health care industry. However, recently medical imaging has expanded the frontiers of AI into areas previously unimaginable. This frontier continues to expand into other areas of medicine, as we discussed, such as clinical practice, translational medicine, and basic biomedical research. Here, I would like to walk you through a non-exhaustive list of some recent AI applications in medicine that demonstrate how AI algorithms can benefit patients, doctors, researchers through making diagnoses, treatments, discoveries, and the practice of medicine faster, more accurate, and more efficient. As I already mentioned, one field that has attracted particular attention for the application of AI in health care is image analysis, including models for interpreting chest X-ray, detecting cancer in mammograms, identifying brain tumors in MRIs, or predicting development of Alzheimer's Disease in PET scans. Here we see how AI is being used in clinical practice for diagnosis of diseases. Let's walk through an example of an AI tool for disease diagnosis using X-rays, which are the most common medical imaging in the world and critical for the diagnosis of common thoracic diseases. The interpretation of these images is time-consuming and there is a shortage of trained radiologist around the globe. Researchers at Stanford Academic Medical Center, recently developed a deep learning algorithm to aid in the terpretation of chest X-rays. The algorithm can detect 14 different pathologies for 10 diseases and can also localize parts of the image most indicative of each pathology. Overall, the algorithm performed just as well as radiologists for one pathology, and it outperformed the experts in some others. Regarding efficiency, radiologist labeled 420 images in 240 minutes on average, and the algorithm labeled them in 1.5 minutes. This work demonstrates the power and efficiency of AI solutions in the medical imaging field and highlights how AI solutions can reduce errors related to human fatigue and improve diagnostics. A rapidly expanding area of AI development under the translational research branch is drug discovery. In another example, researchers from the University of Toronto have developed AI models specifically support vector machine classifiers using various genomic datasets to classify proteins into drug and non-drug targets for different cancer sites. Key classification features were: gene essentiality, mRNA expression, DNA copy number, mutation occurrence, and protein-protein interaction network topology. This study identified 53 novel cancer targets across multiple cancer sites and hundreds of sites specific gene targets. This work demonstrates how AI is being used for translational research and drug discovery moving beyond single drug targets to assays that produce hundreds of potential drug targets for cancer treatment. As another example of AI being used in clinical practice, my group at Stanford has used AI methods for patient risk stratification or population level segmentation. Risk stratification is often difficult at point of care because it relies on the synthesis of heterogeneous information from clinical data, laboratory test results, and imaging. For prostate cancer staging, clinical guidelines recommend a radionucleotide bone scan for patients considered at high risk of reoccurrence of metastatic disease, and recommends no bone scan to be given to patients classified as low risk. Using both structured and unstructured information found in electronic health records, our group developed a tool to stratify prostate cancer patients based on their risk of reoccurrence of metastatic disease as defined by the National Comprehensive Cancer Network. Using a combination of a rule-based method and a convolutional neural network method, we were able to accurately stratify patients based on risk of reoccurrence of metastatic disease. This tool provides a synthesis of heterogeneous information to guide treatment decisions and adherence to clinical guidelines at point of care. This example demonstrates how AI is used in clinical practice for risk stratification. More recently, another group from Stanford, this tool provides a synthesis of heterogeneous information to guide treatment decisions and adherence to clinical guidelines at point of care. This example demonstrates how AI is used in clinical practice for risk stratification up on inpatient encounters and then prospectively validated. The tool was able to partition the entire health information exchange population into three subgroups: low, intermediate, and high risk. These subcategories corresponded to the probability of a hospital readmission as determined by a positive predicted value. In addition, the team performed a time to event analysis and revealed that the higher-risk group re-admitted to the hospital earlier than the lower risk group. Six, high-risk patient subgroup patterns were revealed through unsupervised clustering. The model was deployed into the statewide health information exchange system to identify patient readmission risk upon hospitalization and then daily during hospitalization. This is an example of AI use in clinical practice for risk stratification. AI is also being used in basic biomedical research. Now we will walk through an example of an AI automated experiment. Researchers from the Department of Energy's Brookhaven National Laboratory, have recently developed a fully autonomous experiment using AI. The approach provides automated exploration of multi-dimensional parameter spaces and develops a representative model using the available experimental data. The model development process is refined continuously when new data are collected. The tool detects the regions of greatest measurement uncertainty and then captures more data from that particular region, increasing knowledge with each added measurement. This process reduces the model error and sampling of the parameter space. This is an exciting new area in biomedical research, where AI is being used in basic research for automated experimentation. Finally, as a last example, let's talk about the exciting space of using home videos for autism diagnostics. Currently, standard approaches to diagnose autism spectrum disorder is complicated and can take several hours to complete. However, using machine learning techniques, a team of scientists from Stanford and industry partners collected videos to assess several behavioral features such as eye contact, social smile, etc, and developed eight different machine learning algorithms to diagnose autism. The models performed with greater than 94 percent accuracy compared to expert raders blinded to the diagnosis. Models were prospectively validated on an independent set of videos, also with very high accuracy. This research highlights the potential to use AI for diagnostic purposes and again, demonstrates the accuracy and efficiency of AI solutions. These examples provide just a glimpse of the broad range of AI applications we see emerging in medicine and the opportunities to improve diagnostics, care delivery, access to care and patient outcomes.
