1
00:00:05,050 --> 00:00:07,760
Let us introduce
the error measures

2
00:00:07,760 --> 00:00:11,360
we used in building
the analytics models.

3
00:00:11,360 --> 00:00:16,430
We of course used R squared,
but we also used other measures.

4
00:00:19,550 --> 00:00:22,570
Next measure, the
so-called "penalty error,"

5
00:00:22,570 --> 00:00:26,420
is motivated by the fact
that if you classify

6
00:00:26,420 --> 00:00:29,600
a very high-risk patient
as a low-risk patient,

7
00:00:29,600 --> 00:00:32,299
this is more costly
than the reverse,

8
00:00:32,299 --> 00:00:34,820
namely classifying
a low-risk patient

9
00:00:34,820 --> 00:00:35,980
as a very high-risk patient.

10
00:00:39,410 --> 00:00:44,200
Motivated by this, we
developed a penalty error.

11
00:00:44,200 --> 00:00:48,050
And the idea is to use
asymmetric penalties.

12
00:00:48,050 --> 00:00:51,380
The graph here--
so it's a matrix--

13
00:00:51,380 --> 00:00:54,880
where this is the outcome
and this is the forecast.

14
00:00:54,880 --> 00:01:03,260
For example, whenever we
classify a low-risk patient

15
00:01:03,260 --> 00:01:06,820
as high-risk, we
pay a penalty of 2,

16
00:01:06,820 --> 00:01:12,720
which is a difference of 3 minus
1, the difference in the error.

17
00:01:12,720 --> 00:01:21,250
But inversely, when you
classify a bracket 3 patient

18
00:01:21,250 --> 00:01:24,340
as bracket 1 patient,
this is double.

19
00:01:24,340 --> 00:01:27,150
The cost-- the penalty--
is double the amount.

20
00:01:27,150 --> 00:01:34,340
So you observe that the off
diagonal penalties are double

21
00:01:34,340 --> 00:01:36,860
the corresponding penalties
in the lower diagonal.

22
00:01:39,789 --> 00:01:46,860
To judge the quality of the
analytics models we developed,

23
00:01:46,860 --> 00:01:48,440
we compare it with a baseline.

24
00:01:48,440 --> 00:01:50,100
And the baseline is
to simply predict

25
00:01:50,100 --> 00:01:52,150
that the cost in
the next "period"

26
00:01:52,150 --> 00:01:54,850
will be the cost in
the current period.

27
00:01:54,850 --> 00:01:59,170
We have observed that as far
as identification of brackets

28
00:01:59,170 --> 00:02:03,120
is concerned, the
accuracy was 75%.

29
00:02:03,120 --> 00:02:10,830
So namely, whenever we predict
that the risk is bracket 3--

30
00:02:10,830 --> 00:02:14,410
indeed it is bracket 3--
this happens 75% of the time,

31
00:02:14,410 --> 00:02:16,600
and the penalty error--
the average penalty

32
00:02:16,600 --> 00:02:20,190
error of the
baseline-- was 0.56.