ARTIFICIAL INTELLIGENCE IN Ar Artificial Int Intel elligenc ence e (AI) AI): The ability of machines to HEALTHCARE perform tasks that would normally require human intelligence by giving them the ability to perceive, learn from, abstract, and act using data LAUREN NEAL, PHD PRINCIPAL/DIRECTOR BOOZ ALLEN HAMILTON eHealth Initiative March 2019 This document is confidential and intended solely for the client to whom it is addressed.
CAN MACHINES PERFORM AS WELL AS HUMAN DOCTORS A RECENT NATURE ARTICLE DESCRIBED HOW AI SYSTEMS CAN HELP DOCTORS DIAGNOSE DISEASE. HOW TO DISTINGUISH HYPE VS. REALITY? Booz Allen Hamilton Restricted 1
A BRIEF HISTORY OF AI AI HAS EXISTED SINCE THE 1950S, BUT PROGRESS HAS RECENTLY ACCELERATED Eras of Machine Intelligence Yesterday Today Simple Task Execution Pattern Recognition Tomorrow Machines perform simple, deterministic tasks Contextual Reasoning Machines recognize and act on patterns in in static environments using human knowledge data in static environments using codified as explicit sets of rules and Machines understand context and use it to sophisticated machine learning techniques. programmed into them. make decisions in dynamic environments using sophisticated machine learning techniques. 1950 1960 1970 1980 1990 2000 2010 2020 2030 2040 Accelerating Events of the 2010s 3 2 1 Advances in Computing Power Proliferation of Digital Data New Machine Learning Research Source: Booz Allen analysis, Michael Copeland for Nvidia; Booz Allen Hamilton Restricted 2
HOW MACHINES LEARN THE FIVE “TRIBES” OF MACHINE LEARNING Five approaches to structuring machine learning algorithms 1. Fill in gaps in existing knowledge Technical “Tribe” Origins Motivation Approach Automate the scientific Logic, Inverse 2. Emulate the human brain Symbolists method Philosophy Deduction Reverse engineer the Connectionists Neuroscience human brain via math Backpropagation 3. Simulate evolution over generations model of neurons Replicate the evolution Evolutionary Genetic Evolutionaries of the human brain Biology Programming over generations 4. Systematically reduce uncertainty Test hypotheses to Probabilistic Bayesians Statistics determine the Inference certainty of knowledge Use previous 5. Find similarities between old and new problems / solutions Analogizers Psychology Kernel Machines and extrapolate into new context Source: The Master Algorithm by Pedro Domingos Booz Allen Hamilton Restricted 3
WHAT YOU CAN (AND CAN’T) DO IN THE WORLD OF AI AI IS GOOD AT AUTOMATING SIMPLE TASKS & FINDING/ACTING ON PATTERNS Today, machines can outpace humans on some complex tasks, while a three-year old child can intuitively understand a scenario that even the most advanced AI cannot comprehend. Cannot … Intelligent Machines Can … • Respond to human commands • Speak conversationally about any topic • Drive down a major highway you choose • Select the best treatments for disease • Drive in dense cities or bad weather • Write poems, music, and artwork • Create art that is better than humans’ • Learn human tastes, preferences • Understand human emotion, humor • Outperform humans at strategy games • Invent new games to play • Learn to perform narrow tasks better than • Teach itself new skills independently humans Booz Allen Hamilton Restricted 4
AI TECHNOLOGIES NON-EXHAUSTIVE AI INVESTMENT HAS GIVEN RISE TO EXPANSIVE AI CAPABILITIES AI Era Resulting technologies Use Cases Example Application A software “bot” transposes data Simple task execution • Routine task automation Robotic Process from patient records into an • Process improvement Fully Deployed Automation (RPA) online database • Cognitive automation Core machine learning Software scans patient data to Pattern recognition • Anomaly detection & software identify new indicators of disease Emerging Deployment/ response Pilots • Image/video tagging • Facial recognition A x-ray machine automatically • Biometrics • Scene analysis Computer Vision identifies anomalies in patient scans • Sentiment analysis • Virtual assistants • Language detection Virtual assistants engage with patients Natural Language • Chatbots • Sentiment analysis to ask about symptoms and route them Processing and • Machine translation • Text analysis to the correct care provider Generation • Speech recognition • Report generation • Co-bots A robotic surgeon performs surgery, • Smart manufacturing Cognitive Robotics automatically responding to changes in • Smart logistics a patient’s condition in real time • Fully autonomous A vehicle drives down a crowded city Contextual Reasoning Semantic or “Cognitive” vehicles road, responding to bad weather and In the lab computing obstacles in traffic Source: Booz Allen Analysis Booz Allen Hamilton Restricted 5
AI APPLICATIONS IN HEALTHCARE NON-EXHAUSTIVE AI SHOWS PROMISE IN TACKLING HEALTHCARE CHALLENGES AI can provide solutions that reduce the clerical burden of EHR documentation and augment diagnoses with medical imaging supercomputers. With $30 billion a year flowing into AI research and development, new applications for patient monitoring and disease prediction have the potential to transform patient care. Patient monitoring Imaging & Diagnostics Today, chatbots serve as the first line of support for mental health AI is already being integrated into medical imaging analytics platforms patients, checking in with individuals suffering from depression, to automate volumetric segmentation of lung nodules, detect cardiac monitoring moods, and sharing videos and tools. In the future, function, identify suspected large vessel occlusions, and analyze CT artificial emotional intelligence (AEI) will be used to analyze verbal and perfusion images of the brain using deep learning. non-verbal cues to determine a person’s emotional or psychological state and guide treatment. Speech-enabled EHR Platforms Platforms that provide speech-enabled data entry are being integrated Disease Prediction with EHRs to improve physician-patient interactions. Digital scribes Today, physicians can predict cardiovascular disease based on automatically enter information into the EHR system and virtual AI combined results from blood tests, an EKG and a CT scan. In the assistants analyze conversations between doctors and patients. future, noninvasive scans of the back of the eye will be used to predict the risk of suffering a heart attack or stroke. Beyond heart disease, Clinical Text Processing deep learning will be used to predict Alzheimer’s Disease progression Natural language processing (NLP) extracts relevant medical and detect the location, duration and types of events in EEG time information trapped in EHR clinical notes and supports terminology series to diagnose sleep disorders. mapping. Booz Allen Hamilton Restricted 6
CHALLENGES FOR AI APPLICATIONS IN HEALTHCARE DATA INTENSIVENESS IS A BARRIER FOR ORGANIZATIONS GETTING STARTED WITH AI Lack of talent assets Lack of organized, labeled data AI talent is scarce, and the battle for experts is fierce. Even the most Data is expensive to gather and process, and it is often created for prominent organizations can rarely hold talent for more than a year or billing purposes and not for diagnosis. Data sets also need to be very two. In healthcare, the issue is even more pronounced because AI large, labeled and representative in order to train machine learning experts don’t always understand clinical challenges. algorithms. Ideas to consider: Balance between borrowing, buying and building AI Ideas to consider: Use partnerships and structure infrastructure to talent. For example, partner with academic organizations to borrow capture the data you will need. Consider collecting new data to power world-class talent and invest in programs to upskill in-house staff. your AI efforts. Managing risk Maintaining fairness It’s important to remember that AI systems are still nascent and no AI Machine learning algorithms may work really well for one patient product or platform is truly “off-the-shelf.” All but the most basic group, but results may not be appropriate for others. Without data applications of AI come with a certain level of risk. This is even more that is representative of diverse patient groups, fairness will continue critical when considering healthcare applications. to be a major challenge. Ideas to consider: Start small, then scale. For example, robotic process Ideas to consider: Solicit input from a range of colleagues to ensure a automation (RPA) can be applied relatively easily and quickly to many diversity of perspectives are incorporated into model building efforts. administrative tasks and the cost is also generally low. Make an effort to gather data from diverse patient groups. Source: Artificial Intelligence Primer Booz Allen Hamilton Restricted 11
Recommend
More recommend
