Using The Hadoop Ecosystem to Drive Healthcare Innovation
Download as PPTX, PDF2 likes1,546 views
Aly Sivji's presentation discusses the potential of the Hadoop ecosystem in driving innovation in healthcare through big data and machine learning techniques. It addresses the historical development of electronic health records, the challenges facing the U.S. healthcare system, and how advanced analytics can enhance medical outcomes and reduce costs. Additionally, it explains the functionalities of various Hadoop components, including Apache Tez and Spark, as well as the benefits of using these technologies for data processing and analysis in a healthcare context.
Using The Hadoop Ecosystem to Drive Healthcare Innovation
- 1. Using the Hadoop Ecosystem to Drive Healthcare Innovation Aly Sivji April 25, 2017
- 2. About Me • Aly Sivji – Twitter: @CaiusSivjus – Blog: http://alysivji.github.io • Senior Analyst @ IBM Watson Health – Value-Based Care: Planning Solutions • Grad Student @ Northwestern University – Medical Informatics • Interests: – Technology 🐍 – Data 📈 – Star Trek 🖖🖖
- 3. Overview • Big Data drives most industries
- 4. Overview • What about Healthcare? – Machine Learning • Fraud detection ($65+ billion lost every year) – Wired Article – dataiku - Detecting Medicare Fraud • Preventing unnecessary procedures – Data Mining • Identifying medication prescribed together – Recommender Systems • Finding similar patients
- 5. Overview Healthcare is Different. People who work in healthcare Additional Reading • John Halamka (The Health Care Blog) • Health Catalyst
- 6. Overview • Data Analytics / Data Science – Retrospective versus Predictive • Machine Learning – Types of Algorithms • Healthcare Analytics
- 7. Overview • Apache Hadoop Ecosystem – Big Data framework – Distributed computation on commodity hardware – Demo!
- 8. Road to Electronic Health Records 1920s – Modern record keeping begins 1960s – Dr. Larry Weed introduces problem- oriented medical records 1972 – Regenstrief Institute develops first EMR System 1980s-90s – Siloed adoption by departments & admin 1996 – HIPAA establishes national standards for electronic health records 2004 – President Bush calls for Computerized Health Records
- 9. 2009: EHRs Go Mainstream • HITECH Act passed by President Obama – $25.9 billion to expand Health IT (HIT) adoption • Meaningful Use (MU) program – Incentive payments for using HIT to • Improve quality, safety, efficiency of care • Engage patients • Increase care co-ordination – Goal: MU compliance => better outcomes
- 10. EHR Adoption: Doubled Since 2008 Office-based Physician Electronic Health Record Adoption (2005-2015) Source: Office of the National Coordinator for Health Information Technology. 'Office-based Physician Electronic Health Record Adoption,' Health IT Quick-Stat #50. dashboard.healthit.gov/quickstats/pages/physician-ehr-adoption-trends.php. Dec 2016.
- 11. Health Data Today • Electronic Health Records • Genomic Data ($1000 genome) • Medical Internet of Things (mIoT) • Wearable devices • Bottom Line: Data is growing Big Data = 'Bigger Data' in Healthcare (article)
- 12. Data Analytics • Businesses collect lots of data – IBM: 90% of world’s data created in last 2 years • How can we find hidden patterns in the data and make information actionable? Data Science!
- 13. Types of Analytics • Retrospective Analytics – Summarizing historical activity / performance – Limited scope for making future plans • Better than nothing
- 14. Types of Analytics • Predictive Analytics – Finding patterns (correlations) between historical environment and results – Apply to current environment to make predictions
- 15. Predictive Analytics "Once you have enough data, you start to see patterns. You can then build a model of how these data work. Once you build a model, you can predict.” Michael Wu Chief Scientist, Lithium Technologies
- 17. Machine Learning (ML) “Field of study that gives computers the ability to learn without being explicitly programmed” Arthur Samuel Artificial Intelligence Pioneer
- 18. Machine Learning Algorithms • A probabilistic framework to create models used for predictions • Predictive models are developed iteratively • Models are refined until they converge – i.e. output gets close to a specific value
- 19. Types of ML Algorithms • Unsupervised Learning – Group objects by similar characteristics – Given inputs (X), find label for each observation • Supervised Learning – Given inputs (X) and output (Y) – Find function f that maps X to Y – Given new inputs (Xnew), predict value/label (Ynew)
- 20. Types of Supervised Learning • Regression – Try to predict a value (continuous variable) • Classification – Try to predict a label (discrete variable)
- 21. Analytics in Healthcare “Advanced analytics can be used to improve medical outcomes, increase financial performance, deepen relationships with customers and patients, and drive new medical innovations” Jason Burke Author of Health Analytics
- 22. Healthcare Challenges • US Healthcare spending = $3.4 trillion / year
- 23. Healthcare Challenges • US system wastes $750 billion annually Source: Washington Post (Sept 2012). Retrieved from https://www.washingtonpost.com/news/wonk/wp/2012/09/07/we-spend- 750-billion-on-unnecessary-health-care-two-charts-explain-why/
- 24. Healthcare Challenges • Low quality – To Err is Human Report: • 44,000 - 98,000 deaths to preventable medical errors – Rates poorly when compared to other countries • Last in 2014 Commonwealth Fund survey on: – Quality of care – Access to doctors – Equity
- 25. Solution: Big Data! • Use data analytics and machine learning to improve outcomes & lower costs
- 26. Types of Healthcare Analytics
- 27. Good News • Most of the analytical and software capabilities needed to drive systemic changes in healthcare are already available as: – Commercial software – Open Source solutions 🎉 • Hadoop ecosystem
- 28. Big Data • Characteristics (4 V’s of Big Data) – Volume • Scale of data – Variety • Diversity of data (many sources) – Velocity • Speed of data – Veracity • Certainty of data • 5th V: Value?
- 29. Types of Data • Structured – Highly organized information that fits neatly into a relational database (columns and rows) • Unstructured – Has internal structure, but does not fit into a traditional database (or spreadsheet) – Most data is unstructured (>80%) – Can use Extract-Transform-Load (ETL) Processing to turn unstructured data into structured data
- 30. Apache Hadoop • Set of open source software technology components that form a scalable system we can use to analyze Big Data • Main features: – Distributed storage and processing • Data is too big for a single computer – Runs on commodity hardware – Fault tolerant • Hardware failures are common and handle automatically – Runs in Java Virtual Machine (JVM) environment
- 31. Sample Hadoop Stack Source: Soong, K. (Feb 2016). Big Data Specialization. Retrieved from http://ksoong.org/big-data
- 32. Core Hadoop Components • Yet Another Resource Negotiator (YARN) – “Operating System” for Hadoop – Controls how resources are allocated to different applications and execution engines across cluster
- 33. Core Hadoop Components • Hadoop Distributed File System (HDFS) – Highly scalable storage system Data File
- 34. Core Hadoop Components • Hadoop Distributed File System (HDFS) – Too big to fit on single machine => Partition A B C D
- 35. Core Hadoop Components • Hadoop Distributed File System (HDFS) – Split across multiple machines – Data is protected against hardware failure A B C A D A C D B C D Server 1 Server 2 Server 3 Server 4
- 36. Core Hadoop Components • Hadoop Distributed File System (HDFS) – Server goes down, we can still reconstruct data A B C A D A C D B C D Server 1 Server 2 Server 3 Server 4 🔥
- 37. Core Hadoop Components • Execution Engine – Used when running analytic applications – Distributed data allows us to perform parallel computations – MapReduce execution engine comes bundled with the Hadoop core distribution – Can plug-in different components • Tez, Storm, Spark, etc
- 38. MapReduce Overview Source: Eckroth, J. (n.d.). MapReduce. Retrieved from http://cinf401.artifice.cc/notes/mapreduce.html HDFS HDFS
- 39. MapReduce Example Source: Zhang, X. (Jul 2013). A Simple Example to Demonstrate how does the MapReduce work. Retrieved from http://xiaochongzhang.me/blog/?p=338
- 40. MapReduce Limitations • Lot of read/writes – I/O becomes bottleneck when performing analysis • Machine Learning algorithms are iterative – Many reads and writes cycles before convergence – Slow runtime • There must be a better way!
- 41. Apache Tez • Optimizes workflow to limit number of writes • Less I/O => faster execution
- 42. Apache Storm • Execution engine for real-time streaming applications • Data is analyzed as it is generated BEFORE it is stored
- 43. Apache Spark • In-memory computational engine • Read in data once, subsequent calculations are done in-memory Logistic Regression Runtime
- 44. Other Apache Projects • Apache Hive – SQL interface to data stored in HDFS – Analysts with SQL experience can use Hadoop
- 45. Other Apache Projects • Databases – Apache HBase – Apache Cassandra
- 46. Other Apache Projects • Apache Kafka – Messaging system for streaming data
- 47. Optimal Hadoop Workflow • Depends on what you are trying to do • Data Lake (HDFS) – Storage repository that holds data in raw format – Read into Spark to perform analysis • Use Data Science and Machine Learning algorithms • Demo will walkthrough this workflow
- 49. Dataset • Texas Department of State Health Services – Released State Inpatient / Outpatient data (link) • Inpatient (IP) - 1999 to 2010 • Outpatient (OP) – Q42009 to 2010 – Data is de-identified and made available for free – Tab-delimited text files (for each quarter) • IP data – 450MB base table, 500MB charges • OP data – 750MB base table, 700MB charges
- 50. Spark Background • Java, Scala, Python, and R APIs (docs) • Built around the concept of Resilient Distributed Datasets (RDDs) – Can perform MapReduce on RDD OR – Use the Spark DataFrame abstraction *Recommended*
- 51. Spark DataFrame • Distributed collection of rows and named columns – Think relational database or spreadsheet – Akin to pandas DataFrame or R data.frame # Displays the content of the DataFrame df.show() # # +----+-------+ # | age| name| # +----+-------+ # |null|Michael| # | 30| Andy| # | 19| Justin| # +----+-------+
- 52. Questions? • Slides and code available at https://github.com/alysivji/talks
Editor's Notes
- #3: Before we get to what we’re talking about. I’ll talk about me.
- #4: Data has been making a huge difference in other industries Chase uses machine learning algorithms to flag purchases that could be fraudulent. Last time this happened, I booked my flight using my American Airlines card and booked my hotel and conference on my United card. Chase didn’t know about the flight so it asked for my confirmation. Saves them money for having to pay for fraudulent purchases. Amazon uses data mining to find products purchased together and makes suggestions to increase revenue. Spark was created in Scala and most people who learn Scala do so in order to use Spark in its native language. Amazon doesn’t know this, but it can use data to figure this out. Netflix’s recommendation system finds users who are similar to you and uses their ratings to make predictions for media for you to watch
- #5: Medical fraud dedection could be more robust or similar algorithms can find unnecessary procedures (purchases that do not match my profile) Data mining to suggest medication that is always prescribed together if an order is missing it Recommendation system to find similar patients. Group them by the treatment prescribed, rate their outcomes and use that information to suggest optimal course of action Why is this not widespread in healthcare?
- #6: People who work in healthcare know, healthcare is different. We won’t really go into too many details why, but you can find out more at the links provided. I will spend some time discussing how healthcare has changed and made it easier to facilitate a data revolution
- #7: What do we mean by data revolution? Data is ubiquitous... We’ll explore data science in some depth to understand the basic principles of the field and get a grasp on how we can make our information actionable Bee is Buzzword Bee! I’ll try to include him every time I use a buzzword
- #8: Next we’ll talk about we can use the Hadoop ecosystem to analyze healthcare data
- #9: Is paved with good intentions ;) 1920s [1] Healthcare professionals realized that documenting patient care benefited both providers and patients. Patient records established the details, complications and outcomes of patient care. Once healthcare providers realized that they were better able to treat patients with complete and accurate medical history, documentation became wildly popular. Health records were soon recognized as being critical to the safety and quality of the patient experience. 1960s [2] Charting how we currently know it. First, a patient database is collected. Then use that information to start the diagnosis process. Database is very thorough contains: Family history Prior encounter information Lab results Current health status 1972 [1, 2] There are quite a few cases of electronic record system pilots (thru universities and large healthcare facilities), this is the first major system that was developed. Did not attract many physicians 1980s-90s [1, 2] Computers made their way into hospitals, like they did in every other professional environment, but systems did not speak to each other 1996 HIPAA was passed and national standards for electronic health records was established 2004 [1, 3] In his 2004 State of the Union, President George W Bush calls for computerized health records. Established the Office of the National Coordinator for Health Information Technology. It coordinates nationwide efforts to implement HealthIT and electronic exchange of health information. References [1] http://www.rasmussen.edu/degrees/health-sciences/blog/health-information-management-history/ [2] http://www.nethealth.com/a-history-of-electronic-medical-records-infographic/ [3] https://en.wikipedia.org/wiki/Office_of_the_National_Coordinator_for_Health_Information_Technology
- #10: Meaningful Use provided incentive payments to healthcare providers who could demonstrate they used health information technology in a ‘meaningful way’ to improve quality, engage patients, increase care coordination. Goal is that MU compliance will result in: Better clinical outcomes Improved population health outcomes Increased transparency and efficiency Empowered individuals https://en.wikipedia.org/wiki/Health_Information_Technology_for_Economic_and_Clinical_Health_Act https://www.healthit.gov/providers-professionals/meaningful-use-definition-objectives
- #11: Did it work? Well… it did increase EHR adoption
- #12: * EHR systems have a wealth of data and are collecting more each day * Genomic sequencing costs less than $1000 dollar, I’ve heard about a race to $100 as well * Medical sensors are collecting information at a dizzying pace. One big application is patient sensors in post-acute care environments where patients are hooked up to machines collecting real-time data * People are more concerned about their health than ever before and the consumer wearable industry is growing.
- #13: But we’re getting ahead of ourselves. I need to introduce the topic of data analytics References [1] https://datascience.berkeley.edu/about/what-is-data-science/
- #14: References [1] http://blog.datagravity.com/the-transition-to-predictive-analytics/
- #15: References [1] http://blog.datagravity.com/the-transition-to-predictive-analytics/
- #16: References [1] http://www.informationweek.com/big-data/big-data-analytics/big-data-analytics-descriptive-vs-predictive-vs-prescriptive/d/d-id/1113279
- #17: References [1] https://marketoonist.com/2016/12/predictive-analytics.html
- #18: This leads nicely into the topic of Machine Learning References http://www.ibmbigdatahub.com/blog/how-does-machine-learning-work?cm_mmc=OSocial_Twitter-_-IBM+Analytics_Inbound+Marketing-_-WW_WW-_-B+Yelland+3-20-2017&cm_mmca1=000000VQ&cm_mmca2=10000779&
- #20: References [1] http://machinelearningmastery.com/supervised-and-unsupervised-machine-learning-algorithms/ [2] http://www.ibmbigdatahub.com/blog/how-does-machine-learning-work
- #22: Why is this relevant to us in healthcare?
- #23: Analytics is suited to the specific challenges in healthcare References [1] http://www.pbs.org/newshour/rundown/new-peak-us-health-care-spending-10345-per-person/ [2] http://www.pgpf.org/chart-archive/0006_health-care-oecd
- #25: References [1] https://en.wikipedia.org/wiki/To_Err_is_Human_(report) [2] http://time.com/2888403/u-s-health-care-ranked-worst-in-the-developed-world/
- #27: Healthcare analytics is broad as we can see from this diagram. Lots of areas where a little bit of deliberate data science and machine learning to make a difference
- #28: Worth noting that most of the analytical capabilities needed to drive systemic changes in healthcare are already available in commercial software
- #29: So let’s start talking about Big Data. What is big data? In healthcare, there is a lot of data… each genome is around 200GB of raw data. Lots of different information… clinical, notes, lab information, demographic result data, patient generated data Velocity data... Real time sensors monitoring patients Veracity... How sure are we that the data we get is correct? References [1] http://www.ibmbigdatahub.com/infographic/extracting-business-value-4-vs-big-data
- #30: References [1] https://www.trifacta.com/blog/structured-unstructured-data/ [2] http://sherpasoftware.com/blog/structured-and-unstructured-data-what-is-it/
- #31: How can we deal with all this data? Hadoop Ecosystem!
- #33: References [1] http://www.littlebeelibrary.com/pdfs/Apache_Hadoop.pdf
- #38: Execution engine is used to perform calculations on the underlying data
- #39: The MapReduce engine runs the map step on all nodes in the cluster to produce a set of intermediate output files. It then sorts these intermediate les and then runs a reduce step to take the sorted intermediate les and aggregate the data to get a final result. This process is scalable but relatively slow because of the need to write lots of intermediate les to disk and then read them again.
- #44: The key takeaway from this presentation: Use Spark to do all calculations