How Machines Work in Healthcare.pdf
《How Machines Work in Healthcare.pdf》由会员分享,可在线阅读,更多相关《How Machines Work in Healthcare.pdf(7页珍藏版)》请在得力文库 - 分享文档赚钱的网站上搜索。
1、 IQVIA Real-World Insights How macHineS learn in HealtHcare Machine learning is transforming every facet of healthcare,as computer systems are being taught how to use Big Data to derive insights and support decision making.In this respect,teaching a computer,no less than teaching a child,is to“shape
2、 the future.”Educating a computer is a surprisingly labor-intensive process,requiring massive amounts of data,a nuanced understanding of every data element from every data source,years of trial and error,and extensive domain expertise.The key differentiator in machine learning is not the specific te
3、chnology and science applied;it is in the volume and quality of the instructional material and the knowledge of the instructor.22aDVances in DaTa scienceanDRei sToica,PhD Vice President Systems Development IQVIA Today,were surrounded by computing systems that can learn from experience and handle new
4、 situations.Behind our internet searches,spam filters,online music curation,and virtual assistants,computers are studying away,becoming“smarter”with each interaction we have with them.Machine learning is destined to accelerate the pace of healthcare transformation,as it allows us to extract meaning
5、from otherwise insurmountable volumes of data.It is proving valuable in supporting research and development,identifying populations at risk,improving diagnostics,providing clinical decision support and optimizing sales and marketing.A little understood fact is that a machine learns in much the same
6、way as humans.The ingredients are a scientific model(from simple rules to complex algorithms),information,and a knowledgeable teacher(the domain expert).When these elements come together in the right way,machines are able to perform high-volume automation,recognize patterns,spot anomalies,provide li
7、nkages,offer recommendations,run simulations and make predictions about future outcomes with great reliability.IQVIA was a Big Data company long before the term“Big Data”was coined.Thirty years ago,we had more data than we could effectively move over the internet and had to use elaborate“sneakernets
8、”to transport data.Today,we are in a similar position with machine learning.The term has been over-hyped by vendors that have limited experience with healthcare data.This article is the first in a series examining what it takes to do machine learning in healthcare based on the knowledge of experts t
9、hat have been applying these models and algorithms for decades.We define the entire process from data processing to analytics and the intrinsic interdependencies between the various stages underpinning the quality of results.This article provides additional details about data processing as the found
10、ation for analytics.GooD DaTa hYGiene Sometimes,answering healthcare business questions calls for data of great breadth.Other times,for data of great depth.But in most cases,and especially for business critical decisions,the data must be clean.Thats why most data mining systems that claim they work
11、on“dirty data”have in fact an intensive data-cleansing step prior to data processing.Its worth reviewing the three basic steps involved in data cleansing and processing:bridging,coding and linking.These steps not only prepare the data,but they are the foundation for quality machine learning in proce
12、ssing and analytics stages.All healthcare records contain multiple references to entities(such as diagnoses,products,physicians,procedures,outlets and companies,etc.)And there can be thousands of attributes linked to each entity.For example,a medical encounter can have hundreds of attributes,includi
13、ng details on the procedures,imaging,notes,etc.In some cases,there are standard codes by which these entities can be referenced,such as the National Drug Code(NDC),a universal product identifier for human drugs in the U.S.Where standard codes such as this exist,they must be assigned to the entity in
14、 the data record and subsequently validated.This assignment is called bridging.If the entity does not have a standard code,a unique one must be created as a reference in a process called coding.To prepare data for bridging and coding,simple rules are first introduced to the computer.For example,one
15、basic rule might be to remove all extra blank spaces in names.Machine learning requires human,healthcare knowledgeMachine learning is destined to accelerate the pace of healthcare transformation,as it allows us to extract meaning from otherwise insurmountable volumes of datacontinued on next pageAcc
16、essPoint Volume 7 Issue 14 23ADVANCeS IN DATA SCIeNCeThen,with greater exposure to more data and situations,complex machine-learning algorithms(such as neural networks,score engines,Ngrams,random forests,Bayesian networks,genetic algorithms and many others)begin reasoning about data attributes in ev
17、ery individual data stream.At this stage,machines can differentiate,for example,between a doctor who has just changed her name after marriage vs.another doctor with the same last name that just graduated medical school and started practicing in the same city.Experience shows that complex machine-lea
18、rning inferences used to bridge and code data on pharmaceutical packaging and dosage,doctor addresses,distribution outlets and volumes,patients and medical procedures and hundreds of other attributes must be highly specific even down to the individual data stream or data supplier.Achieving this leve
19、l of specificity requires deep,detailed knowledge of the field.only someone with a history of working with a given supplier would know,for instance,that promotions for an individual pharmacy store are coded as one record while promotions at the chain-store level are coded as multiple transactions.Th
20、is example is just one simple rule that IQVIAs massive machine-learning infrastructure has learned over time.Currently there are hundreds to,in some cases,thousands of rules for each of our 800,000 data suppliers around the world.Once an entity in a record is properly bridged or coded,it can be link
21、ed to records in other data sets to allow for cross-referencing.Billions of healthcare records are generated and exchanged between different parties in the healthcare industry each year,and accurate bridging and coding must take place with every data exchange to maintain the quality and usefulness o
22、f the data.“Dirty”data can have significant implications for the decisions that pharmaceutical companies make from research and development to commercial planning.The privacy laws of different countries add complexity to the challenge of linking records.Where laws allow de-identified data to be coll
23、ected,the de-identification algorithms have to be both secure(irreversible)as well as versatile to allow linkage for longitudinal patient analytics for example.For de-identification,homogenous systems are key to data security and linkage.It is less likely that data de-identified with different stand
24、ards can be linked,or if it is,that it will not affect the security of the data.a case in PoinTMachine learning is changing healthcare in real time.IQVIA built a decision-support system using machine learning to help sponsors manage physician selection in clinical trials a task fundamental to trial
25、success.experts in the therapeutic area defined multi-dimensional models to express all of the study protocol details.Data scientists with complementary expertise worked as part of the same team to define matching multi-dimensional models to express all physician prescribing/treatment patterns and h
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- How Machines Work in Healthcare
限制150内