[Skolkovo Robotics V] Application of AI in Healthcare
Dr. Hans-Aloys Wischmann
Innovation Program Manager, Philips Research
Филипс Инновационные лаборатории РУС
Application of
AI in Healthcare
Robotics V, Skolkovo
April 21, 2017
A stroke event in 2025
95-year old male, living alone, is preparing breakfast …
t=0 Suffers a stroke, collapses to the ground
15 sec Fall detected, chat initiated, slurred speech observed, ambulance dispatched
15 min Patient is in self-driving ambulance, paramedics focus on patient, telemetry
30 min Arrival in hospital, IV line in ED by HaemoBot, CT started -> inconclusive
40 min MRI started
60 min AI-auto-segmentation of all structures in head/brain/vessels -> ischemic stroke
65 min tPA administered (< 4 hours)
70 min Nano-Robot inserted into bloodstream to mechanically eliminate clot
90 min Patient walks out of hospital, takes a taxi to work
A stroke event in 2025
95-year old male, living alone, is preparing breakfast …
t=0 Suffers a stroke, collapses to the ground
15 sec Fall detected, chat initiated, slurred speech observed, ambulance dispatched
15 min Patient is in self-driving ambulance, paramedics focus on patient, telemetry
30 min Arrival in hospital, IV line in ED by HaemoBot, CT started -> inconclusive
40 min MRI started
60 min AI-auto-segmentation of all structures in head/brain/vessels -> ischemic stroke
65 min tPA administered (< 4 hours)
X 70 min Nano-Robot inserted into bloodstream to mechanically eliminate clot
70 min Robot-Thrombectomy (< 6 hours), short ICU recovery, next-day discharge
StanfordPhilips
The global healthcare challenge
>1/3 of us will be diagnosed with
cancer in their lifetime
500+ million people suffer from
respiratory diseases
400 million people worldwide have diabetes
~1 billion adults with hypertension
Digital is driving exponential growth of health data
Personal health tracking Medical imaging Patient Monitoring
Home monitoring Medication adherence Pathology Quantification Genomics Analytics
We are gaining deeper, denser and more longitudinal insights than ever before
125 years of innovation
For 125 years, we have been improving people’s lives with a steady flow of innovations.
1930
2017
20171960
Philips, a strong position in data science and AI
958 million
radiology studies under management
9.7 million IoT devices
connected to the Internet via HealthSuite Cloud
4 million
sleep therapy patients supported
23 petabytes
of imaging study data
managed for healthcare
providers
275
million
patients
tracked with our
patient monitors
last year
145 billion
images managed
We already served over
7 million seniors
with our wearable Lifeline
service
35 million patients
Supported with population health management
Acute care
Using Deep Learning
and Machine Learning
techniques to extract
meaning from noisy
data.
Early detection of
deterioration through
predictive algorithms.
Precision medicine
AI can supplement the skills of human
radiologists by identifying subtle changes in
imaging scans more quickly, potentially leading to
earlier and more accurate diagnoses.
AI challenges in Healthcare
Access to large
quantities of data with
appropriate patient
privacy protection
Need for curated clean
databases
Need to avoid AI bias
to prevalent diseases
and population in
training data sets
Effective combination
of data and knowledge
driven learning
Understanding of the
(local) clinical context
Need for review cycles
with clinicians and
policy makers / society
Fragmented Healthcare
IT infrastructure
Need to explain AI
“thinking” behind a
diagnosis or therapy
decision to patients
Key takeaways
AI can enable a health care that is more predictive and more precise, fully integrated
More accessible to more people around the globe via networked connected care
AI requires careful co-creation
High-tech should always be applied to enable and support high-touch
AI will support the future care professional
Ambient data collection will be
increasingly paramount
10% Access to care
20% Genetics
20% Environment
50% Healthy behaviorsAnalyze how socio-economic,
behavioral, genetic and
clinical factors correlate and
impact our health.
IoT, sensors, cloud enable
digital environments that are
sensitive and responsive to
the presence of people.
Editor's Notes
Made for Robotics conference
Intend to work on AI-Auto-Segmentation of Head/Brain… in Skolkovo with Skoltech
28% of IV line fails on first attempt
StrokeAssociation 2017: Many people miss this key brain-saving treatment because they don't arrive at the hospital in time for alteplase treatment, which is why it's so important to identify a stroke and seek treatment immediately for the best possible chance at a full recovery.
https://www.eurekalert.org/pub_releases/2013-08/mgh-uot082113.php: Among all patients who were admitted to the participating hospitals for ischemic stroke, usage of tPA increased from 4 percent in 2003 to 7 percent in 2011. In patients who arrived early and were without medical conditions that would prevent safe use of the drug, tPA administration increased from 43 percent to 77 percent. Since the researchers only analyzed data for patients arriving within 2 hours of symptom onset, the increased tPA usage was not due to expansion of the time window.
Made for Robotics conference
Intend to work on AI-Auto-Segmentation of Head/Brain… in Skolkovo with Skoltech
28% of IV line fails on first attempt
StrokeAssociation 2017: Many people miss this key brain-saving treatment because they don't arrive at the hospital in time for alteplase treatment, which is why it's so important to identify a stroke and seek treatment immediately for the best possible chance at a full recovery.
https://www.eurekalert.org/pub_releases/2013-08/mgh-uot082113.php: Among all patients who were admitted to the participating hospitals for ischemic stroke, usage of tPA increased from 4 percent in 2003 to 7 percent in 2011. In patients who arrived early and were without medical conditions that would prevent safe use of the drug, tPA administration increased from 43 percent to 77 percent. Since the researchers only analyzed data for patients arriving within 2 hours of symptom onset, the increased tPA usage was not due to expansion of the time window.
We have a major challenge in healthcare with a growing population of people with chronic conditions and the associated rising costs (global numbers: 400M diabetes, more than 500 million people suffer from respiratory diseases, an estimated 1 billion adults with hypertension).
The demand for care will grow.
While the number of healthcare professionals will decrease …
According to the United Nations, the world’s population is expected to increase by one billion people by 2025. Of that billion, 300 million will be people aged 65 or older, as life expectancy around the globe continues to rise.
By 2050 one fifth of the world’s population will be 60+ (2 billion people), of that one twenty fifth will be 80+ (390 million people)
The Brookings Institute estimates 65% of the global population will be middle class by 2030 which will add to the increase of people living with chronic lifestyle diseases and able to afford access to quality healthcare.
The world will be short of 12.9 million health-care workers by 2035; today, that figure already stands at around 7.2 million.
In 2015, U.S. health care spending increased 5.8 percent to reach $3.2 trillion.
The EU spends around 10% of its GDP on healthcare and faces substantial challenges with increasing chronic disease, graying populations and escalating costs. Between 70% and 80% of European healthcare costs are spent on chronic care, amounting to €700bn in the EU. Chronic diseases account for over 86% of deaths in the EU.
Only 3% of healthcare budgets of the 28 EU Member States are spent on prevention, whereas 80% of cardiovascular diseases, 90% of diabetes 2 and 50% of cancers are preventable.
We simply will not have a workforce to deliver care in the traditional model to a rapidly growing, and economically developing world population. This reality, and changing financial pressures today strongly drive the interest in predictive preventative population health management and large scale analytics.
As our world has become digital the amount and granularity of the medical information we have available has increased exponentially.
Today, all medical images are digital and with some of Philips customers generating more than 2 million images a week the amount of imaging data is growing exponentially.
Our monitors are digital. Vast amounts of patient monitoring information can be streamed in realtime into advanced cockpits that offer contextual information to give a comprehensive patient picture for mission critical clinical decision support in ICUs.
In our pockets we carry smartphones with computing power millions of times more powerful than all of NASA’s combined computing in 1969 when the first Apollo went to the moon.
To give an indication of the vast amounts of data we are talking about, look at Philips alone ….
Many of our customers are generating 2 million medical images …. a week!
We manage 145 billion images adding up to 23 petabytes of imaging study data
We monitor 275 million people yearly in ICUs, General Wards, and in their homes
We work on Genomics. sequence whole genome sequence of a person amounts to a staggering 200Gb of data
With all this data we are able to gain deeper, denser and more longitudinal insights than ever before based on the information we gather over time through for instance monitoring or by integration of data from various sources, both historical and current.
Philips is using data science, data analytics and AI to improve its solutions for health management and healthcare.
Let me briefly explain to you where we come from.
Philips has a 125 year history in working with consumers and 100 years of experience in healthcare.
We have unique insights in consumer behavior but also the clinical context.
We have served 8 million elderly people with our Lifeline emergency response solutions
4 million people with sleep disorders use our sleep solutions
40% of all patient monitors in the world are from Philips
Through our long-standing and current leadership positions we have the unique opportunity to use intelligent algorithms to mine enormous stores of structured and unstructured data for innovative insights. As an example we are mining data from millions of subscribers of Lifeline services and learn from their behavior and situations to further improve the solutions we develop. We can develop new and advanced algorithms based on our historical and current data from the users of our technology.
Of course we perform analytics and AI in compliance with all data privacy and protection laws and standards, respecting patient rights.
We have been doing so for more than 15 years already. Philips has been active in data science and Artificial Intelligence since early 2000.
Today we have 60% of our people working in R&D work on Healthcare Informatics and a big part of them are working on research and the development of meaningful application of Artificial Intelligence in health and healthcare for patients and care professionals and society at large. We are tying together our singularly unique experience in consumer technology, clinical technology and informatics to provide new solutions to patients and providers that will accelerate positive change in the healthcare delivery model.
ACUTE CARE - Clinical Decision Support and Predictive Analytics. AI in the general ward and emergency department|
Acute care clinicians are challenged to assimilate large quantities of multimodal data in order to make diagnosis and therapy decisions.
The potential of guardian EWS in combination with continuous monitoring via wearable biosensors.
Applying data mining and machine learning methodology to critical care large databases to create algorithms for early alerts, risk stratification, therapy decision support.
Early detection of deterioration: predictive algorithms
Therapy decision support: track patient response to therapy, search existing database for similar patient cohort to derive therapy recommendations
Resource management: rank patient population by acuity, estimate length of stay
With IntelliVue Guardian with EWS we already see ICU transfers from General Ward go down by 68%.
Analyzing vast amounts of combined data points on the ICU in order to deliver on preventing alarms
The Acute Care episode is a high care intensity moment in the Continuum of Providing Care, and the stakes are by definition, high as well.
A patient’s pre-event status can provide rich context to the current acute episode to be used for
better outcomes.
Likewise, a better understanding of the post-event status can speed up recovery, prevent readmissions
For both the pre and post intervention period… we can use…
Deep Learning and Machine Learning techniques to extract meaning from noisy data, enabling deterioration detection, clinical pathways guidance, and evidenced-based medicine.
Advanced data visualization to facilitate insights, consultation and decision making by Physicians, Nurses
The past decade we have been spearheading the creation of ICU Datasets of 100-10k patients coming from a single hospital. We were typically talking about ~10’s of parameters and simple algorithms.
As we are moving towards care networks ICU Datasets today cover 50k-2M patients, multiple hospitals. We are talking structured data: all vital signs, medications, lab values, etc., ~100’s of parameters
As we move forward datasets will span Home, ED, ICU and more, 2M-10M patients, dozens of hospitals, structured and unstructured data: free text notes, claims, medications, etc., 1000’s of parameters. Algorithms will be personalized to the patient. Considering patient history and likely therapy responses. Hospitals will monitor and optimize care in real time, bringing evidence based medicine to the bedside and the insights we will be providing will be increasingly impactful.
PRECISION MEDICINE
a) Imaging analytics to see what the human eye can’t seeWe are dedicated to supplementing human capabilities to improve patient outcomes – making providers more efficient and more effective in the process.
Applying AI in radiology for anatomical context, to learn user preferences, and to automate measurements of organs and lesions.
We have also demonstrated the use of AI to improve the disease screening, from detecting tuberculosis in chest x-rays (Sidra Middle East), to screening for lung cancer from high risk patients using CT.
We are using AI to spot very subtle changes over time in scans. Machine learning can supplement the skills of human radiologists by identifying subtle changes in imaging scans more quickly, potentially leading to earlier and more accurate diagnoses. Comparative databases. Comparing a scan with the normal, based on measurements on large-scale population level.
We have also used AI in digital pathology scans to uncover cancer types, to understand the relationship between the local microenvironment and disease, and to help select therapies and get information on projected outcomes.Philips and PathAI, a company that develops artificial intelligence technology for pathology, are collaborating with the aim to develop solutions that improve the precision and accuracy of routine diagnosis of cancer and other diseases. The partnership aims to build deep learning applications in computational pathology enabling this form of artificial intelligence to be applied to massive pathology data sets to better inform diagnostic and treatment decisions. The initial focus of this effort is on developing applications to automatically detect and quantify cancerous lesions in breast cancer tissue. Another tangible example in this field is our new Illumeo technology that is using AI to help radiologists work with medical images. Its anatomical intelligence is able to understand what is shown on the screen and automatically offers the right set of tools for the diagnosis.
We have also utilized AI to combine large sets of structured and unstructured data in a holistic practical ‘Patient Briefing’ that includes the patient problem list, laboratory results, prior radiology reports, imaging orders or scanned documents (such as handwritten referral letters from GPs and referring specialists) obtained from health information systems like the Electronic Medical Record (EMR) or Radiology Information Systems (RIS). It integrates and organizes this data and helps radiologists to bridge the stages of diagnosis to treatment to follow-up, while being able to rely on a current, comprehensive patient picture.
Through Natural Language Processing technology we are extracting meaning from unstructured data with to aim to add it to the body of knowledge. We are using NLP for Illumeo.
The semantic labeling of the image datasets in Illumeo uses NLP techniques, and to a more limited extend we also apply NLP on radiology reports to present the summary section first. We have work in progress towards future releases on more advanced use of NLP to extract meaningful insights from unstructured data. In addition to the immediate impact on clinical care, we will be working with our provider partners such as MD Anderson Cancer Center, using artificial intelligence, to understand when imaging is necessary, as in the case of following cancer patients, and when it is not. In addition to improving outcomes, this will lower the cost of delivering care.
One of the more interesting breakthroughs in the past years has been in genomics, understanding “the code of life”.
A decade or so ago, we passed an interesting threshold – a time at which the human mind could not comprehend what it is made up of, the human cell. The way the human body works and responds still often eludes our best scientists and doctors and AI will help them better understand what drives health and diseases.
AI applied to genomics will help us scan staggering amount of genomics data to find patterns in groups of patients that will help find correlations between personal drivers and drivers on patient cohort level (contextual drivers such as social economic, regional, comorbidities, therapy history): big data analytics on patient groups.
Genomics combined with the full patient context (Lab results, pathology, imaging, family history, therapy response in the past, etc) will help us to understand why one therapy will work for person A but not for person B helping us to enable personalized treatment with more predictable outcomes. For this we need advance healthcare informatics and analytics.
POPULATION HEALTH - Predictive Analytics will help enable preventative health management on population level
As the population ages, so does a desire to age in place when possible, and to maximize not only disease management, but quality of life as we do so.
AI driven Ambulatory Care Programs for elderly people living with multiple chronic conditions. Development of algorithms to make these programs more preventative. First step in this direction is Lifeline with CareSage where we developed algorithms based on 1000s of patient years of data enabling an intelligent technology that can predict the need for emergency transport to the hospital 30 days in advance.Recent study results demonstrated an insightful correlation between chronic conditions and falls risk. Philips’ researchers retrospectively analyzed the records of 145,000 seniors equipped with a standard Philips Lifeline medical alert service or a medical alert service with AutoAlert (automatic fall detection) between January 2012 and June 2014. Data showed seniors with chronic conditions fell and required emergency transport up to 54 percent more often, compared to their peers with no chronic conditions. Additionally, the analysis revealed that seniors with physical conditions not typically tied to frailty, including COPD and diabetes, also were shown to fall more often.
The data shows that seniors fell more often and needed hospital transport when reporting the following:
Cognitive impairment by 54 percent;
COPD by 42 percent;
Diabetes by 30 percent; and
Heart condition by 29 percent.
We are already seeing some amazing results coming out of our home monitoring programs. An example is the program at Banner Health in Phoenix AZ in the USA. By using connected devices and integrating personal health measurement data with hospital data, applying analytics, to prioritize those patients in most need of intervention we were able to reduce hospitalizations by 49.5 percent. Costs were reduced by 35%.
This is likely only further to improve as we are able to integrate more data points and larger amounts of data and apply future even more advanced algorithms based on large scale population analytics.
We can enable a health care that is more predictive and more precise, fully integrated
More accessible to more people around the globe via networked connected care
It requires a careful co-creation
High-tech should always be applied with the aim to enable high-touch
AI will support the future care professional
Healthcare AI could be a way for humans to be increasingly interdependent and collaborative on the basis of the one set of variables – unlike things such as religion, gender and race – that we all share, our biology, our susceptibility to disease, and our desire to be well.
AI is will save your life someday
Future AI will enable consumer-driven, digital propositions that support self-management
We know that asking individuals to do many things outside the context of the normal lives is not optimal, in a healthcare context or otherwise, so ambient data collection will be increasingly paramount.
Wireless technologies, smart environments and AI will play a vital role in healthcare delivery, by optimizing patient pathways and supporting clinical decision making. Innovations such as the Internet of Things (IoT) have paved the way for something known as “Ambient Intelligence”. This refers to digital environments that are sensitive and responsive to the presence of people: devices that recognize users, configure themselves to the user and add to the understanding about that user. For instance, a monitor that recognizes the patient and automatically adapts the right settings and starts streaming vital signs, that can be combined with clinical history of the patient to create the right insights in the condition of that patient.
Ambient Intelligence refers to technologies that seamlessly weave into a consumer’s life to augment their capabilities and support management of chronic conditions. Today’s solutions for monitoring falls, vital signs, medication compliance and immersive experiences to improve health behaviors will become integrated. They will be proving their efficacy, while sustaining a patient’s independence and quality of life.
On top of this, by enabling the virtual visits by a care giver, these innovations have the potential to reduce costs for the healthcare system, minimize unnecessary patient travel, allow individuals to continue to be productive at home and at work, and positively impact those close to the patient.
Alongside these innovations and technological enablers, such as IoT and AI, we’ll see the industry increasingly take into account behavioral and social determinants of health (the social and economic environment, the physical environment, and the person’s individual characteristics and behaviors, access to healthcare services).
Artificial intelligence coupled with new emerging technologies such as IoT, sensors and cloud computing will enable a new era of ambient intelligence and anytime, anywhere health management.
It will make it possible to aggregate and make sense of the health data from millions of consumers and analyze how socio-economic, behavioral, genetic and clinical factors correlate and impact their health.
With this wealth of information that these new technologies make available, we can now gain full patient profiles that offer contextual information from a range of sources (mobile and connected devices, medical records, lab results etc) and develop preventative personalized health programs and targeted therapies aiming for better outcomes.