WEBVTT

00:00:09.490 --> 00:00:13.220
Let us examine what we
learned in this lecture.

00:00:13.220 --> 00:00:17.790
We built an expert-trained
model by a physician

00:00:17.790 --> 00:00:21.920
that can accurately identify
diabetic patients receiving

00:00:21.920 --> 00:00:23.410
low quality care.

00:00:23.410 --> 00:00:26.930
We have observed that the out
of sample accuracy of the model

00:00:26.930 --> 00:00:28.730
was 78%.

00:00:28.730 --> 00:00:32.600
But most importantly, the
model identifies most patients

00:00:32.600 --> 00:00:36.520
receiving poor care, which
is the major objective

00:00:36.520 --> 00:00:37.220
in the study.

00:00:39.780 --> 00:00:42.980
Logistic regression
models provide

00:00:42.980 --> 00:00:48.970
probabilities of somebody
receiving poor quality care.

00:00:48.970 --> 00:00:53.630
These probabilities can be
used to prioritize patients

00:00:53.630 --> 00:01:00.920
for intervention, a particularly
useful outcome from the study.

00:01:00.920 --> 00:01:05.860
While the accuracy is
reasonably high, 78%,

00:01:05.860 --> 00:01:08.120
it can be, of course,
further improved.

00:01:08.120 --> 00:01:13.530
In that respect, I expect that
electronic medical records, not

00:01:13.530 --> 00:01:16.400
only claims, could
be used in the future

00:01:16.400 --> 00:01:19.590
to enhance the predictive
capability of such models.

00:01:19.590 --> 00:01:25.100
So a model like the
one we built can

00:01:25.100 --> 00:01:30.090
be used to analyze literally
millions of records.

00:01:30.090 --> 00:01:35.620
Whereas a human can only
accurately analyze rather small

00:01:35.620 --> 00:01:37.440
amounts of information.

00:01:37.440 --> 00:01:43.039
So clearly such a model
allows significantly larger

00:01:43.039 --> 00:01:43.539
scalability.

00:01:46.800 --> 00:01:51.289
Of course models do not
replace expert judgement.

00:01:51.289 --> 00:01:59.880
However, models provide a way
to translate expert judgement

00:01:59.880 --> 00:02:04.910
to a reproducible, testable
prediction methodology that

00:02:04.910 --> 00:02:10.340
has significantly higher
scalability, as we discussed.

00:02:10.340 --> 00:02:14.490
And of course experts can
continuously improve and refine

00:02:14.490 --> 00:02:17.590
the model, as we have
seen in this lecture.

00:02:21.420 --> 00:02:24.680
Finally, and quite
importantly, models

00:02:24.680 --> 00:02:28.050
can integrate assessments
of multiple experts

00:02:28.050 --> 00:02:31.590
into one final, unbiased,
and unemotional prediction.

00:02:34.180 --> 00:02:38.020
And such methods of combining
assessments and combining

00:02:38.020 --> 00:02:41.880
models is a tool
that we will use

00:02:41.880 --> 00:02:44.550
later in the class
on multiple occasions

00:02:44.550 --> 00:02:51.079
as a way of enhancing and
improving quantitative models.