WEBVTT

00:00:04.490 --> 00:00:07.240
Claims data offers an expansive
view of the patients health

00:00:07.240 --> 00:00:08.230
history.

00:00:08.230 --> 00:00:10.850
Specifically, claims
data include information

00:00:10.850 --> 00:00:14.710
on demographics, medical
history, and medications.

00:00:14.710 --> 00:00:17.240
They offer insights
regarding a patient's risk.

00:00:17.240 --> 00:00:21.710
And as I will demonstrate,
may reveal indicative signals

00:00:21.710 --> 00:00:23.050
and patterns.

00:00:23.050 --> 00:00:25.980
We'll use health
insurance claims

00:00:25.980 --> 00:00:29.690
filed for about 7,000
members from January, 2000

00:00:29.690 --> 00:00:34.620
until November, 2007.

00:00:34.620 --> 00:00:37.980
We concentrated on members
with the four main attributes.

00:00:37.980 --> 00:00:41.200
At least five claims with
coronary artery disease

00:00:41.200 --> 00:00:44.720
diagnosis, at least five
claims with hypertension

00:00:44.720 --> 00:00:48.480
diagnostic codes, at least
100 total medical claims,

00:00:48.480 --> 00:00:51.420
at least five pharmacy
claims, and data

00:00:51.420 --> 00:00:53.740
from at least five years.

00:00:53.740 --> 00:00:57.020
These selections yield
patients with a high risk

00:00:57.020 --> 00:00:58.030
of heart attack.

00:00:58.030 --> 00:01:00.740
And a reasonably rich medical
history with continuous

00:01:00.740 --> 00:01:01.240
coverage.

00:01:04.819 --> 00:01:08.360
Let us discuss how we've
aggregated this data.

00:01:08.360 --> 00:01:11.930
The resulting data sets
includes about 20 million health

00:01:11.930 --> 00:01:14.520
insurance entries, including
individual, medical,

00:01:14.520 --> 00:01:16.150
and pharmaceutical records.

00:01:16.150 --> 00:01:19.600
Diagnosis, procedures,
and drug codes in the data

00:01:19.600 --> 00:01:22.640
set comprised tens of
thousands of attributes.

00:01:22.640 --> 00:01:25.460
The codes were
aggregated into groups.

00:01:25.460 --> 00:01:29.500
218 diagnosis groups,
180 procedure groups,

00:01:29.500 --> 00:01:33.150
538 drug groups.

00:01:33.150 --> 00:01:35.350
46 diagnosis groups
were considered

00:01:35.350 --> 00:01:38.720
by clinicians as possible risk
factors for heart attacks.

00:01:43.240 --> 00:01:46.759
Let us discuss how we
view the data over time.

00:01:46.759 --> 00:01:49.940
It is important in this study
to view the medical records

00:01:49.940 --> 00:01:53.100
chronologically, and to
represent a patient's diagnosis

00:01:53.100 --> 00:01:54.770
profile over time.

00:01:54.770 --> 00:01:59.970
So we record the cost and number
of medical claims and hospital

00:01:59.970 --> 00:02:02.390
visits by a diagnosis.

00:02:02.390 --> 00:02:06.440
All the observations we have
span over five years of data.

00:02:06.440 --> 00:02:11.890
They were split into 21
periods, each 90 days in length.

00:02:11.890 --> 00:02:15.110
We examine nine months
of diagnostic history,

00:02:15.110 --> 00:02:18.440
leading up to heart attack
or no heart attack event,

00:02:18.440 --> 00:02:23.980
and align the data to make
observations date-independent,

00:02:23.980 --> 00:02:26.910
while preserving
the order of events.

00:02:26.910 --> 00:02:29.950
We recorded the diagnostic
history in three periods.

00:02:29.950 --> 00:02:33.230
Zero to three months
before the event,

00:02:33.230 --> 00:02:35.090
three to six months
before the event,

00:02:35.090 --> 00:02:37.600
and six to nine months
before the event.

00:02:40.280 --> 00:02:42.940
What was a target variable
we're trying to predict?

00:02:42.940 --> 00:02:44.940
The target prediction
variable is the occurrence

00:02:44.940 --> 00:02:46.350
of a heart attack.

00:02:46.350 --> 00:02:49.730
We define this from a
combination of several claims.

00:02:49.730 --> 00:02:52.060
Namely, diagnosis
of a heart attack,

00:02:52.060 --> 00:02:54.700
alongside a trip to
the emergency room,

00:02:54.700 --> 00:02:58.000
followed by subsequent
hospitalization.

00:02:58.000 --> 00:03:00.510
Only considering
heart attack diagnosis

00:03:00.510 --> 00:03:03.340
that are associated with the
visits to an emergency room,

00:03:03.340 --> 00:03:05.810
and following
hospitalization helps

00:03:05.810 --> 00:03:09.050
ensure that the target outcome
is in fact a heart attack

00:03:09.050 --> 00:03:10.100
event.

00:03:10.100 --> 00:03:12.190
The target variable is binary.

00:03:12.190 --> 00:03:14.670
It is denoted by
plus 1 or minus 1

00:03:14.670 --> 00:03:16.790
for the occurrence or
non-occurrence of a heart

00:03:16.790 --> 00:03:19.790
attack in the targeted
period of 90 days.

00:03:22.720 --> 00:03:24.090
How's the data organized?

00:03:24.090 --> 00:03:26.690
There were 147 variables.

00:03:26.690 --> 00:03:29.650
Variable one is the patient's
identification number,

00:03:29.650 --> 00:03:32.030
and variable two is
the patient's gender.

00:03:32.030 --> 00:03:35.160
There were variables related
to the diagnoses group

00:03:35.160 --> 00:03:38.410
counts nine, six, and three
months before the heart attack

00:03:38.410 --> 00:03:39.660
target period.

00:03:39.660 --> 00:03:44.950
There were variables related
to the total course nine, six,

00:03:44.950 --> 00:03:46.910
and three months before
the heart attack target

00:03:46.910 --> 00:03:51.630
period, and the final
variable for 147,

00:03:51.630 --> 00:03:54.870
includes the classification of
whether the event was a heart

00:03:54.870 --> 00:03:56.020
attack or not.

00:03:58.860 --> 00:04:02.940
Cost of medical care is a good
summary of a person's health.

00:04:02.940 --> 00:04:08.110
In our database, the total cost
of medical care in the three 90

00:04:08.110 --> 00:04:15.160
day periods preceding the heart
attack target event ranged from

00:04:15.160 --> 00:04:20.760
$0.00 to $636,000
and approximately 70%

00:04:20.760 --> 00:04:24.880
of the overall cost were
generated by only 11%

00:04:24.880 --> 00:04:26.360
of the population.

00:04:26.360 --> 00:04:28.920
This means that the
highest patients

00:04:28.920 --> 00:04:32.460
with high medical expenses
are a very small proportion

00:04:32.460 --> 00:04:35.810
of the data, and could
skew our final results.

00:04:35.810 --> 00:04:39.700
According to the American
Medical Association, only 10%

00:04:39.700 --> 00:04:42.640
of individuals have
projected medical expenses

00:04:42.640 --> 00:04:45.790
of approximately
$10,000 or greater

00:04:45.790 --> 00:04:48.960
per year, which is more
than four times greater

00:04:48.960 --> 00:04:52.470
than the average projected
medical expenses of 2,400

00:04:52.470 --> 00:04:53.880
per year.

00:04:53.880 --> 00:04:57.090
To lessen the effects of
these high-cost outliers,

00:04:57.090 --> 00:04:59.960
we divided the data into
different cost buckets,

00:04:59.960 --> 00:05:03.770
based on the findings of the
American Medical Association.

00:05:03.770 --> 00:05:06.140
We did not want to
have too many cost bins

00:05:06.140 --> 00:05:08.840
because the size
of the data set.

00:05:08.840 --> 00:05:12.140
The table in the
slide gives a summary

00:05:12.140 --> 00:05:13.990
of the cost bucket partitions.

00:05:13.990 --> 00:05:17.550
Patients with expenses over
$10,000 in the nine month

00:05:17.550 --> 00:05:20.980
period were allocated
to cost bucket 3.

00:05:20.980 --> 00:05:23.840
Patients with less
than 2,000 in expenses

00:05:23.840 --> 00:05:25.910
were allocated to cost bucket 1.

00:05:25.910 --> 00:05:30.170
And the remaining patients with
costs between 2,000 and 10,000

00:05:30.170 --> 00:05:31.850
to cost bucket 2.

00:05:31.850 --> 00:05:40.360
Please note that the majority
of patients, 4,400 out of 6,500,

00:05:40.360 --> 00:05:44.590
or 67.5% of all patients
fell into the first bucket

00:05:44.590 --> 00:05:46.470
of low expenses.