Analytics in healthcare has been called everything from the
only path to reduced costs & improved outcomes to a distraction that will
cost substantial money & energy. As always (well, almost always), the
actual fact of the matter lies somewhere in between these extreme views. In
this post, I’ll look at how analytics are currently being used in healthcare
& how they could be used, what the advantages & impediments are &
make some predictions – I am supposed to be a futurist after all.
First, a definition – What do I mean by analytics? I mean
the systematic analysis of data focused on answering a specific question or set
of questions. Analytics is not report generation, nor is it an IT function, a
software package or a technical methodology. It is a way of thinking about an
organization’s goals that makes use of highly focused analysis of data, not
just “big data”, but relevant data. Any analysis that is done must be aligned
with an organization’s goals & strategies. Otherwise it may not produce
results that lead to actions relevant to the organization.
An analytic adoption model for healthcare organizations,
developed by a group of industry experts[1]
has been proposed to allow the evaluation of data warehousing & analytics
efforts. It consists of the following levels:
1. Integrated Enterprise Data Warehouse
2. Standardized Vocabularies & Patient Registries
3. Automated Internal Reporting
4. Automated External Reporting
5. Clinical Effectiveness & Accountable Care
6. Per Case Payment & the Triple Aim
7. Per Capita Payment & Predictive Analysis
8. Per Unit of Health Payment & Prescriptive Analysis
This framework emphasizes the functions associated with
successful data warehousing as well as the relationship of payment with various
dimensions of providing healthcare: per case, per capita & per unit of
health. Is this an interesting way of looking at the evolution of analytics.
Looking at part of the current landscape will help answer that.
I say part of because so much work is currently being done
that it’s difficult to get the picture as a whole. What we can get is some
dimensions or areas of focus & some examples in each area (each area has
many, many more examples)[2]:
- Point-of-Care Recommendations – generally large amount of clinical data analyzed to provide best match to individual patient characterization for delivery of best practice diagnosis & treatment recommendations, not rule-based, but empirical
- Mayo Clinic – 5M clinical records (approximately 15-25PB) are analyzed to provide best practice treatment for individual patients in real time
- Beth Israel Deaconess Medical Center – 2M+ clinical records are analyzed to provide best practice treatment for individual patients in real time
- Kaiser Permanente – 9M clinical records over 10 years (approximately 45PB) are analyzed to provide best practice treatment for individual patients in real time, natural language query system in use
- Partners Healthcare – combined clinical, operational & financial data, best practice recommendations made at time-of-encounter
- Outcome & Population Characterization
- Intermountain Healthcare (partnered with Deloitte) – 90M patient records, two analytic applications developed, Outcome Miner: derives factors that contribute to outcome at the individual level & Population Miner: derives relationship(s) between treatments & outcomes at the population level
- McKinsey – “next 5% analysis”, analysis of 30M commercial claims to determine “micro-segments of patient population that will allow the identification of the top 6% of patients with regard to cost, assignment of care managers
- Predictive Modeling
- ExpressScripts – 1.5B prescriptions/year, use of predictive modeling to determine which patients are most likely to not use prescription as indicated, suggests proactive interventions
- Research
- Mt. Sinai Medical Center (partnered with Ayasdi) – used unique analytic method developed by Ayasdi (topological analysis) to evaluate the entire e. coli genome (including 1M DNA variants) to determine bacteria’s response to different antibiotics
- Operations Optimization
- Oregon Health & Science University – PAR (periodic automatic replenishment) levels for 4000 infusion pumps established & pump utilization & inventory tracked & optimized
Point-of-Care recommendations are far & away the most
numerous applications while research & Operations Optimization appear to be
the least. Point-of-Care systems seem to fall at Level 5 (clinical
effectiveness & accountable care) on the adoption model. Very few efforts
seem, at this time, to fall in the levels above Level 5, which implies that
work on price & cost has not yet been emphasized.
So, if point-of-care, analysis of outcomes & predictive
modeling of various kinds are the current areas of analytic focus, what areas
might be interesting & productive that are not being emphasized today? I
have been surveying a variety of healthcare organizations (informally) by
talking to people at conferences, meetings etc., & this is what I’ve been
hearing.
- Can various forms of trend analysis be combined with geo-locational data to provide insight into very local conditions, for instance: Are specific diagnoses concentrated locally & if so are they associated with specific clinical characterizations?
- Can trend analysis of changes in population served be combined with larger scale demographic data, for instance: Are large-scale demographic trends driving trends in numbers of patients, ethnic grouping of patients etc.? Are larger-scale demographic trends going to influence whether healthcare organizations & specific locations should be invested in, for instance: In an area where the overall population is decreasing rapidly, should clinics, health centers of hospitals remain open? What factors, other than demographic trends, should influence these decisions?
- Is there a relationship between cost of care & cost of outcome on a per patient, per provider &/or per location basis?
- Can data on service utilization & demographics be used to model service utilization trends for planning purposes?
- Are there bottlenecks in clinical & operational workflows that affect quality of care & outcome (it’s not clear that the data for this analysis is generally available, although several business process modeling methodologies could be used to address the issue)?
These are just a few of the many areas of inquiry &
specific inquiries people have talked to me about. Only the largest healthcare
organizations I’ve spoken with (Partners Healthcare, Vanderbilt Medical Center
etc.) have focused on point-of-care recommendations as an analytic goal. This
is, in part, because only these organizations have access to the ultra-large
clinical data sets that are optimal for this type of analysis. Even where data
aggregates, such as data warehouses, have been developed by medium & small
sized organizations, the organizations have tended to focus on specific
operational & financial questions as sustainability is their biggest issue.
This generalization, like all such generalizations, must be taken for what it
is worth – a generalization from not very much data.
What, then, are the impediments to the use of analytics as I
have been describing it. There are the obvious ones of lack of resources, lack
of expertise, lack of experienced personnel etc., but for small-to-medium size
healthcare organizations, the largest impediment I have seen is the lack of an
approach to providing appropriate data for analysis. Many organizations are
using conventional data warehouse & extract techniques that at a minimum require
semantic & syntactic normalization as well as transformation &
standardization of the data in order to have meaningful results. I have seen,
reviewed & participated in a number of projects over the last year in which
results were not usable because data were aggregated without this work being
done (in many cases by an outside contractor hired by the healthcare
organization). I have also participated in a project where the definition of
core elements like “encounter”, “outcome” & “provider” could not be determined
from the data & could not be agreed upon by the project participants[3]. As we say, “Garbage in, Garbage out.[4]”
Even those organizations that are using more contemporary
analytic methods, such as Hadoop-based analytic stacks, often have trouble with
use of data for lack of experience & expertise. This will improve as more
organizations move to these methods & more people are trained in their use.
The second major impediment that I have seen while working
with healthcare organizations is a lack of understanding of what analytics is
& can do – this is especially true in reference to the difference between
analytics & reporting. All healthcare organizations do reporting from their
clinical & practice management data: what quality measures scored highest
in the practice? Lowest? What departments had the highest numbers of patients
etc. Analytics is different in that we
are trying to explore less obvious, & in some cases non-obvious &
unintuitive relationships in the data, often in very large data sets. I have
had many people that I work with on this ask me if “analytics” can improve
their reporting of quality measures. The answer is yes, if you are looking for
underlying factors affecting performance, but not if you a simply trying to get
better results on quality performance from your data. This is mainly an
experience & education issue.
Of course, these issues are a result of many of the obvious
impediments: lack of resources, expertise & so on…
So where are we on healthcare analytics? It seems clear (to
me at least) that the use of analytic techniques to explore clinical &
operational data will become more & more important in the next several to
the extent that if a healthcare organization is not developing this expertise
& using it to try to optimize clinical & operational efforts, that
organization will fall behind in the effort to meet the Triple Aim of improving
the experience of care, improving the health of populations & lowering the
cost of care per capita. Organizations can use pattern matching in ultra-large
data sets to provide improved diagnosis & treatment planning at
point-of-care which is of core importance to the patient, but they will also
have to begin to explore the relationships between cost & provision of care
for both individuals & populations in order to begin to lower such costs,
& begin to do predictive modeling & trend analysis in order to be able
to optimize utilization of scarce resources & to sustain their operations.
The inclusion of methodologies from other disciplines such as business process
modeling for workflow optimization, other modeling & simulation techniques
for optimization of efforts such as: CPOE (inventory & ordering models),
pharmacy utilization, other operations optimization & even analysis of
social media data (to extract information of use in clinical & operational
workflows) will be essential. We are just at the beginning of the analytics
effort & “letting a thousand flowers bloom[5]”
is the best way to move toward a consolidation & consensus of what works.
Five years from now, healthcare organizations will routinely be analyzing data
for the continuous improvement of clinical & operational efforts & actual
meaningful use will include such analysis as a core part of what healthcare
organizations do. If it doesn’t, we’ll still be stuck in our current morass of
huge amounts of data, but little insight or wisdom about how to provide care
& control costs.
Next up: More on social media in healthcare (I may be
getting too obsessed with this…)
[3] It is interesting (?) that
the only outcome they could unanimously agree on was the death of the patient.
[4] Apparently first used in a syndicated newspaper story
about early computerization efforts at the Internal Revenue Service in April,
1963; although my favorite (earlier) example comes from Charles Babbage who in
1864 wrote,” On two
occasions I have been asked, "Pray, Mr. Babbage, if you put into the
machine wrong figures, will the right answers come out?" ... I am not able
rightly to apprehend the kind of confusion of ideas that could provoke such a
question.” Babbage, C. 1864. Passages from the Life of a Philosopher. Longman
& Co. London. P.67.
[5] Although most people think
that when Mao ZeDong initiated the Hundred Flowers Campaign in 1956, it was to
allow dissidents to express themselves so they could be identified & dealt with.
