WEBVTT
1
00:00:10.349 --> 00:00:14.429
Hello data standard community. I'm your
host, Shanea Weldon, and this is
2
00:00:14.509 --> 00:00:18.309
the data standard audio experience. We
have a great guest today on our podcast.
3
00:00:18.429 --> 00:00:22.699
We have Jose Lumpinko, as senior
director for data science and machine learning
4
00:00:22.739 --> 00:00:27.940
at West healthy, has python software
development skills and high level applied mathematics,
5
00:00:28.100 --> 00:00:32.539
including deep learning in your own networks, Ma shine learning and optimization. Welcome
6
00:00:32.579 --> 00:00:35.219
to the show, Jose. All
right, thank you. Thanks for the
7
00:00:35.299 --> 00:00:38.649
opportunity. Appreciate you having you.
Welcome. I appreciate you willing to be
8
00:00:38.770 --> 00:00:42.049
on here to sort off the pod
Gat. Yeah, the start off.
9
00:00:42.369 --> 00:00:46.250
Can you please just go ahead intus
yourself? Okay, my name is Dr
10
00:00:46.289 --> 00:00:50.079
Jsse and Pinko. I am the
senior director for data science at the West
11
00:00:50.159 --> 00:00:54.799
Health Institute, which is a nonprofit
medical research organization. I'm also a lecturer
12
00:00:54.960 --> 00:00:58.600
at the University of California San Diego
or machine learning and data science, both
13
00:00:58.679 --> 00:01:02.719
the undergraduate and degrade graduate programs,
and I'm also the author of three books
14
00:01:02.759 --> 00:01:07.670
using Python for scientific applications, including
machine learning, statistics and signal processing.
15
00:01:07.870 --> 00:01:12.709
Wows, so impressive. So work
at West Health. What services does less
16
00:01:12.750 --> 00:01:19.299
health divide? So West town is
a what's called a nonprofit medical research organization.
17
00:01:19.459 --> 00:01:26.140
So it's mission really is to do
research into integrated healthcare models that improve
18
00:01:26.340 --> 00:01:30.659
a health outcomes and, a particular, reduced healthcare costs that also meets the
19
00:01:30.939 --> 00:01:34.650
also including non medical needs for the
seniors in the US. So the whole
20
00:01:34.730 --> 00:01:41.209
idea is to come up with better
models and influence sort of policy development that
21
00:01:41.370 --> 00:01:45.769
influences the wellbeing of seniors us.
What kind of data do you work with?
22
00:01:46.209 --> 00:01:49.560
So we have a wide range of
partnerships with other institutions, other hospitals
23
00:01:49.599 --> 00:01:55.040
or healthcare providers or other institutions that
provide services like, for example, meal
24
00:01:55.079 --> 00:01:57.640
delivery services to the seniors in the
US. So, as a result,
25
00:01:57.680 --> 00:02:00.870
we have a wide variety of data. That sort of depends on what the
26
00:02:00.950 --> 00:02:04.750
motive other research is. So we
have a lot of, for example,
27
00:02:04.829 --> 00:02:08.349
claims data. These are like after
you go you see us your for your
28
00:02:08.430 --> 00:02:13.150
doctor, for example, there's usually
a payment right of paying the insurance company
29
00:02:13.189 --> 00:02:15.620
makes. So we have lots of
that kind of data. We have also
30
00:02:15.780 --> 00:02:21.740
lots of state and federal other data
that the government collects data on poverty,
31
00:02:21.780 --> 00:02:25.300
for example. We have a geographic
data where different people are located. Your
32
00:02:25.300 --> 00:02:30.449
were seniors are located, concentrations of
poor people are located in us, and
33
00:02:30.610 --> 00:02:35.129
then we have also health people called
encounter data, sort of like you know,
34
00:02:35.610 --> 00:02:38.009
what happened, whatever, whatever,
like when you go to the doctor's
35
00:02:38.050 --> 00:02:40.569
office and they might always be typing
in the stuff. You know, that
36
00:02:40.650 --> 00:02:45.479
kind of sort of encounter based data
we have. So we have a mix
37
00:02:45.800 --> 00:02:50.879
of different data sets and a lot
of the challenge is really combining or developing
38
00:02:50.919 --> 00:02:54.719
new data sets that address specific research. It okay, right in you say
39
00:02:55.199 --> 00:03:00.349
so. I know there's a huge
misconception about data flowing freely without defects.
40
00:03:00.710 --> 00:03:06.909
The reality it's actually dollated by defects. So how do you handle this in
41
00:03:07.030 --> 00:03:10.110
the industry? So I think one
thing I've noticed in particular when I see
42
00:03:10.110 --> 00:03:15.020
students, it's they are in the
impression that high quality datas or something mouse
43
00:03:15.060 --> 00:03:19.860
click away, I can use click
and download this data, and then the
44
00:03:20.620 --> 00:03:23.300
value in the data science product is
really in, you know, the machine
45
00:03:23.300 --> 00:03:29.409
learning model or that. So in
their minds it's really the value add comes
46
00:03:29.449 --> 00:03:32.889
from constructing this model. In reality, especially in the area that I work
47
00:03:32.930 --> 00:03:37.569
in, the data science, value
add actually begins at the point where you're
48
00:03:37.610 --> 00:03:42.319
determining what data should be collected that
will lead to a valuable product. So
49
00:03:42.759 --> 00:03:46.879
the reduction or combination and imputation of
existing data sets is really where the data
50
00:03:46.960 --> 00:03:52.680
science activity is most important, because
this is where you're essentially putting a down
51
00:03:52.680 --> 00:03:55.430
payment on what the resulting value of
the product will be. So yeah,
52
00:03:57.110 --> 00:04:01.669
you can't escape garbaging garbage out.
So it's really the refinement and the understanding
53
00:04:01.710 --> 00:04:05.110
of what needs to be collected to
go to a specific goal, to to
54
00:04:05.590 --> 00:04:09.710
wind up with a valuable data science
product of the end. So and and
55
00:04:09.830 --> 00:04:14.740
that involves cleaning and talking to people
who have domain knowledge specific issues that drive
56
00:04:14.819 --> 00:04:19.579
towards the motives what the collaborator wants. I heard you mention garbage and garbage
57
00:04:19.620 --> 00:04:23.060
out. I know this is one
of your terms of your professor tell you
58
00:04:23.540 --> 00:04:27.329
you also have a term called health
data paradox. So what is that play?
59
00:04:27.410 --> 00:04:30.610
Is that me? So, when
you're talking about sort of the cost
60
00:04:30.649 --> 00:04:33.290
of health cares, what happens is
so machine running models consume a lot of
61
00:04:33.329 --> 00:04:35.689
data, right, so if you're
talking about health data, they're going to
62
00:04:35.730 --> 00:04:39.279
want to consume a lot of health
date. So where is there plenty of
63
00:04:39.360 --> 00:04:43.040
health data? Well, there's plenty
of health data among people who are chronically
64
00:04:43.079 --> 00:04:46.079
all. Right, if you're admitted
to the hospital, believe me, you're
65
00:04:46.079 --> 00:04:49.160
going to generate a lot of high
quality data just by the fact that you're
66
00:04:49.160 --> 00:04:54.389
constantly a mounder under observation. We've
being treated right. So from a cost
67
00:04:54.430 --> 00:04:57.750
standpoint, the ship has sailed right. From a cost standpoint, it's a
68
00:04:57.790 --> 00:05:00.750
little late to worry about that when
you're already you know, I see you
69
00:05:01.110 --> 00:05:04.709
getting all these stuff. So yes, you're machine learning models are going to
70
00:05:04.709 --> 00:05:11.220
be heavily instrumented for the problem is
there that the selection by its drives the
71
00:05:11.579 --> 00:05:15.339
models to essentially predict essentially sick people
will get sicker. Right, and that's
72
00:05:15.420 --> 00:05:18.660
not an interesting question to ask.
When you're dealing about, for example,
73
00:05:18.699 --> 00:05:21.689
of cause of your you're interested in
the ones who have yet to be sick,
74
00:05:21.730 --> 00:05:26.089
right, potential, and those are
the ones for which you have very
75
00:05:26.089 --> 00:05:28.209
little data, right, because if
you're not sick, you're obviously not going
76
00:05:28.250 --> 00:05:30.610
to go with the doctrine constantly be
measured for this man. So that's really
77
00:05:30.689 --> 00:05:38.000
the paradox. Where there is strong
data collect is amongst people for whom costs
78
00:05:38.040 --> 00:05:41.439
have already happened. Right. So
that as part of the de collect but
79
00:05:41.600 --> 00:05:45.439
that's not, in particular, where
your most interested if you're interested in reducing
80
00:05:45.480 --> 00:05:49.000
the cost of healthcare and making healthcare
more available to, in particular, seniors
81
00:05:49.000 --> 00:05:53.589
in the US. Okay, thanks, side description, and that's also as
82
00:05:53.629 --> 00:05:56.949
the couple to talk about, but
on experience plaining. Now, what is
83
00:05:56.990 --> 00:06:01.870
the situation in which the risks of
models are calculated based on individual outcomes and
84
00:06:02.069 --> 00:06:05.660
not aggregate outcomes? It's an excellent
question. So if you're dealing in a
85
00:06:05.779 --> 00:06:13.100
situation where the results of your data
science activity can dramatically impact an individuated person,
86
00:06:13.259 --> 00:06:16.019
right, you're not infecting people,
you're impacting persons. Right. So,
87
00:06:16.459 --> 00:06:23.529
for example, if my healthcare model
says that Mrs Jones should not yet
88
00:06:23.889 --> 00:06:28.970
like a particular surgery or should not
get a particular intervention, the potential downside
89
00:06:29.089 --> 00:06:30.529
from that could really be severe.
So, of the machine learning model,
90
00:06:30.529 --> 00:06:33.759
it's only rating on sort of the
minimum means squared error, I. which
91
00:06:33.759 --> 00:06:40.759
is sort of an average effect.
Right, then you've missed the individuated risk.
92
00:06:41.000 --> 00:06:44.560
That can happen, not just in
medical but in military applications or financial
93
00:06:44.639 --> 00:06:48.990
applications. If somebody doesn't get alone
or has some severe financial downside from the
94
00:06:49.029 --> 00:06:51.230
output. If your mode you can
go around saying Oh, well, on
95
00:06:51.310 --> 00:06:58.990
average worked right, because the severe
outcomes aren't captured sort of in the metric
96
00:06:59.069 --> 00:07:02.819
that's being used to to collect and
qualify the model. Right. So it's
97
00:07:02.819 --> 00:07:08.019
very seldom, and I'm actually never
seen it in practice where the out the
98
00:07:08.220 --> 00:07:13.579
mathematical objective of a machine running model
is sufficient to satisfy the organization seeds.
99
00:07:13.660 --> 00:07:17.129
I've actually never seen this. It's
always that the value that is a combination
100
00:07:17.370 --> 00:07:23.009
of the objectives of the machine learning
model but as their influenced by the other
101
00:07:23.129 --> 00:07:25.449
needs of the organization. So,
for example, we have a lot of
102
00:07:25.529 --> 00:07:29.250
privacy and security needs. Right,
no one's going to care about the outcome
103
00:07:29.250 --> 00:07:31.360
of this particular about how great it
is, how low the minimum means squared
104
00:07:31.399 --> 00:07:36.199
error is, if the other needs
of the organization unfolded into into how value
105
00:07:36.319 --> 00:07:41.439
is is, you know, recognized
by the model that we're going in a
106
00:07:41.519 --> 00:07:45.470
data and a little issues. I
guess what is the biggest challenge that you
107
00:07:45.589 --> 00:07:49.110
face and your industry? So I
think anybody dealing with sensitive data. I
108
00:07:49.230 --> 00:07:53.910
think security and privacy. As you
know, with all the constant news reports
109
00:07:53.949 --> 00:07:58.620
of everybody you know being held hostage
for data, being hijacked, etcetera.
110
00:07:58.740 --> 00:08:03.459
So anybody who deals with any kind
of secure information, in particular health conformation,
111
00:08:03.699 --> 00:08:09.579
big challenges for Security and privacy.
So maintaining trusting relationships with the other
112
00:08:09.779 --> 00:08:11.300
collaborators that you have not just that
you say, Oh, you know,
113
00:08:11.379 --> 00:08:15.730
we're secure and we're good. You
have to be in a situation where you
114
00:08:15.769 --> 00:08:20.490
have policies and procedures in place that
can demonstrate artifacts that that effect, that
115
00:08:20.649 --> 00:08:24.529
can head off problems before they happen. So I think security and privacy are
116
00:08:24.930 --> 00:08:28.040
big issues and are going to become
bigger issues as more data becomes a or
117
00:08:28.120 --> 00:08:33.440
the value of data in so far
as it recognizes issues of privacy and security
118
00:08:33.720 --> 00:08:37.159
become more common. And then the
other thing that we have is data collection
119
00:08:37.279 --> 00:08:41.399
is not for your cheap, in
particular in this area. So the ask
120
00:08:41.789 --> 00:08:46.429
for collecting data that directly influence as
a stated goal is something yet to be
121
00:08:46.590 --> 00:08:52.750
cognizant of. So collecting data actually
that's that drives towards a specifical isn't so
122
00:08:52.870 --> 00:08:56.340
easy when you're when you're dealing with
with healthcare dat so and the world I
123
00:08:56.419 --> 00:09:00.980
work in, those two most dominant. I think the security and privacy one
124
00:09:01.100 --> 00:09:05.019
is going to become a bigger deal
when you're dealing with other other kinds of
125
00:09:05.100 --> 00:09:07.820
industries. I just think the healthcare
one is sort of, you know,
126
00:09:07.299 --> 00:09:11.970
healthcare of financial military. I mean
they're already up into that, but I
127
00:09:13.009 --> 00:09:16.690
think that's gonna become a bigger deal
in future. Security, privacy and collecting
128
00:09:16.730 --> 00:09:20.169
data. Okay, so we get
the data standard. Are Building a community
129
00:09:20.210 --> 00:09:24.360
where data enthusiasts to collaborate things,
to connect. And so what can we
130
00:09:24.440 --> 00:09:30.000
do for you? So, Um, there's a great overview of the research
131
00:09:30.159 --> 00:09:33.360
that we do here at the institute
on the West Health Dot org site.
132
00:09:33.440 --> 00:09:35.840
So that's a lot. That's a
great place to keep up sort of on
133
00:09:35.879 --> 00:09:37.590
the overall. There's a lot more
to it actually, I was able to
134
00:09:37.629 --> 00:09:41.950
get into we also have a Github
site where we have where we support a
135
00:09:41.990 --> 00:09:45.230
number of open source products, in
particular visualization products. So you can go
136
00:09:45.389 --> 00:09:50.110
to gethubcom slash the West health you
can see some of the open collaborations we
137
00:09:50.230 --> 00:09:54.299
have regarding visual relation tools in particular, and other things that we that we
138
00:09:54.379 --> 00:09:58.100
do for the community. And then, if you're interested, there's also my
139
00:09:58.179 --> 00:10:03.259
own get up side. Get UPCOM
and Pinko has all the Jupiter notebooks corresponding
140
00:10:03.259 --> 00:10:05.419
to the books I've written if you
want to connect. So there's plenty of
141
00:10:05.500 --> 00:10:09.649
ways you can collaborate with us on
a technical level and research level. So
142
00:10:09.210 --> 00:10:13.570
we're happy to talk to anybody's interested. Okay, thanks for the staying the
143
00:10:13.610 --> 00:10:18.210
websites and then also work in viewers. Find you at that same website you
144
00:10:18.289 --> 00:10:22.919
just listed. I get home or
Christian, so it's a githubcom and then
145
00:10:22.440 --> 00:10:28.799
the last name it's you and pausing
Peter I and is never g CEO,
146
00:10:28.519 --> 00:10:31.720
and you'll find all the other material
I have on there. Involved in a
147
00:10:31.879 --> 00:10:35.190
number of other open source projects as
well. So yeah, we're happy to
148
00:10:35.350 --> 00:10:39.470
collaborate in where we have were always
we have papers we put out of the
149
00:10:39.549 --> 00:10:45.110
various python conferences. So we're big
python shop here. So well, thank
150
00:10:45.149 --> 00:10:48.470
you and thanks for, yeah,
beings allowing to be kind of the community.
151
00:10:48.950 --> 00:10:54.100
So you can find us at the
data standard www dot data standard ioh
152
00:10:54.460 --> 00:10:58.779
this episode was also sponsored by Pandio
and they're innovating the text base and you
153
00:10:58.820 --> 00:11:01.899
can find them at CANOCOM. But
back to is a thank you so much.
154
00:11:03.220 --> 00:11:05.690
You're insided for being here. Today, oh, thanks, thanks for
155
00:11:05.690 --> 00:11:07.169
your opportunity. I really pushing
1
00:00:10.349 --> 00:00:14.429
Hello data standard community. I'm your
host, Shanea Weldon, and this is
2
00:00:14.509 --> 00:00:18.309
the data standard audio experience. We
have a great guest today on our podcast.
3
00:00:18.429 --> 00:00:22.699
We have Jose Lumpinko, as senior
director for data science and machine learning
4
00:00:22.739 --> 00:00:27.940
at West healthy, has python software
development skills and high level applied mathematics,
5
00:00:28.100 --> 00:00:32.539
including deep learning in your own networks, Ma shine learning and optimization. Welcome
6
00:00:32.579 --> 00:00:35.219
to the show, Jose. All
right, thank you. Thanks for the
7
00:00:35.299 --> 00:00:38.649
opportunity. Appreciate you having you.
Welcome. I appreciate you willing to be
8
00:00:38.770 --> 00:00:42.049
on here to sort off the pod
Gat. Yeah, the start off.
9
00:00:42.369 --> 00:00:46.250
Can you please just go ahead intus
yourself? Okay, my name is Dr
10
00:00:46.289 --> 00:00:50.079
Jsse and Pinko. I am the
senior director for data science at the West
11
00:00:50.159 --> 00:00:54.799
Health Institute, which is a nonprofit
medical research organization. I'm also a lecturer
12
00:00:54.960 --> 00:00:58.600
at the University of California San Diego
or machine learning and data science, both
13
00:00:58.679 --> 00:01:02.719
the undergraduate and degrade graduate programs,
and I'm also the author of three books
14
00:01:02.759 --> 00:01:07.670
using Python for scientific applications, including
machine learning, statistics and signal processing.
15
00:01:07.870 --> 00:01:12.709
Wows, so impressive. So work
at West Health. What services does less
16
00:01:12.750 --> 00:01:19.299
health divide? So West town is
a what's called a nonprofit medical research organization.
17
00:01:19.459 --> 00:01:26.140
So it's mission really is to do
research into integrated healthcare models that improve
18
00:01:26.340 --> 00:01:30.659
a health outcomes and, a particular, reduced healthcare costs that also meets the
19
00:01:30.939 --> 00:01:34.650
also including non medical needs for the
seniors in the US. So the whole
20
00:01:34.730 --> 00:01:41.209
idea is to come up with better
models and influence sort of policy development that
21
00:01:41.370 --> 00:01:45.769
influences the wellbeing of seniors us.
What kind of data do you work with?
22
00:01:46.209 --> 00:01:49.560
So we have a wide range of
partnerships with other institutions, other hospitals
23
00:01:49.599 --> 00:01:55.040
or healthcare providers or other institutions that
provide services like, for example, meal
24
00:01:55.079 --> 00:01:57.640
delivery services to the seniors in the
US. So, as a result,
25
00:01:57.680 --> 00:02:00.870
we have a wide variety of data. That sort of depends on what the
26
00:02:00.950 --> 00:02:04.750
motive other research is. So we
have a lot of, for example,
27
00:02:04.829 --> 00:02:08.349
claims data. These are like after
you go you see us your for your
28
00:02:08.430 --> 00:02:13.150
doctor, for example, there's usually
a payment right of paying the insurance company
29
00:02:13.189 --> 00:02:15.620
makes. So we have lots of
that kind of data. We have also
30
00:02:15.780 --> 00:02:21.740
lots of state and federal other data
that the government collects data on poverty,
31
00:02:21.780 --> 00:02:25.300
for example. We have a geographic
data where different people are located. Your
32
00:02:25.300 --> 00:02:30.449
were seniors are located, concentrations of
poor people are located in us, and
33
00:02:30.610 --> 00:02:35.129
then we have also health people called
encounter data, sort of like you know,
34
00:02:35.610 --> 00:02:38.009
what happened, whatever, whatever,
like when you go to the doctor's
35
00:02:38.050 --> 00:02:40.569
office and they might always be typing
in the stuff. You know, that
36
00:02:40.650 --> 00:02:45.479
kind of sort of encounter based data
we have. So we have a mix
37
00:02:45.800 --> 00:02:50.879
of different data sets and a lot
of the challenge is really combining or developing
38
00:02:50.919 --> 00:02:54.719
new data sets that address specific research. It okay, right in you say
39
00:02:55.199 --> 00:03:00.349
so. I know there's a huge
misconception about data flowing freely without defects.
40
00:03:00.710 --> 00:03:06.909
The reality it's actually dollated by defects. So how do you handle this in
41
00:03:07.030 --> 00:03:10.110
the industry? So I think one
thing I've noticed in particular when I see
42
00:03:10.110 --> 00:03:15.020
students, it's they are in the
impression that high quality datas or something mouse
43
00:03:15.060 --> 00:03:19.860
click away, I can use click
and download this data, and then the
44
00:03:20.620 --> 00:03:23.300
value in the data science product is
really in, you know, the machine
45
00:03:23.300 --> 00:03:29.409
learning model or that. So in
their minds it's really the value add comes
46
00:03:29.449 --> 00:03:32.889
from constructing this model. In reality, especially in the area that I work
47
00:03:32.930 --> 00:03:37.569
in, the data science, value
add actually begins at the point where you're
48
00:03:37.610 --> 00:03:42.319
determining what data should be collected that
will lead to a valuable product. So
49
00:03:42.759 --> 00:03:46.879
the reduction or combination and imputation of
existing data sets is really where the data
50
00:03:46.960 --> 00:03:52.680
science activity is most important, because
this is where you're essentially putting a down
51
00:03:52.680 --> 00:03:55.430
payment on what the resulting value of
the product will be. So yeah,
52
00:03:57.110 --> 00:04:01.669
you can't escape garbaging garbage out.
So it's really the refinement and the understanding
53
00:04:01.710 --> 00:04:05.110
of what needs to be collected to
go to a specific goal, to to
54
00:04:05.590 --> 00:04:09.710
wind up with a valuable data science
product of the end. So and and
55
00:04:09.830 --> 00:04:14.740
that involves cleaning and talking to people
who have domain knowledge specific issues that drive
56
00:04:14.819 --> 00:04:19.579
towards the motives what the collaborator wants. I heard you mention garbage and garbage
57
00:04:19.620 --> 00:04:23.060
out. I know this is one
of your terms of your professor tell you
58
00:04:23.540 --> 00:04:27.329
you also have a term called health
data paradox. So what is that play?
59
00:04:27.410 --> 00:04:30.610
Is that me? So, when
you're talking about sort of the cost
60
00:04:30.649 --> 00:04:33.290
of health cares, what happens is
so machine running models consume a lot of
61
00:04:33.329 --> 00:04:35.689
data, right, so if you're
talking about health data, they're going to
62
00:04:35.730 --> 00:04:39.279
want to consume a lot of health
date. So where is there plenty of
63
00:04:39.360 --> 00:04:43.040
health data? Well, there's plenty
of health data among people who are chronically
64
00:04:43.079 --> 00:04:46.079
all. Right, if you're admitted
to the hospital, believe me, you're
65
00:04:46.079 --> 00:04:49.160
going to generate a lot of high
quality data just by the fact that you're
66
00:04:49.160 --> 00:04:54.389
constantly a mounder under observation. We've
being treated right. So from a cost
67
00:04:54.430 --> 00:04:57.750
standpoint, the ship has sailed right. From a cost standpoint, it's a
68
00:04:57.790 --> 00:05:00.750
little late to worry about that when
you're already you know, I see you
69
00:05:01.110 --> 00:05:04.709
getting all these stuff. So yes, you're machine learning models are going to
70
00:05:04.709 --> 00:05:11.220
be heavily instrumented for the problem is
there that the selection by its drives the
71
00:05:11.579 --> 00:05:15.339
models to essentially predict essentially sick people
will get sicker. Right, and that's
72
00:05:15.420 --> 00:05:18.660
not an interesting question to ask.
When you're dealing about, for example,
73
00:05:18.699 --> 00:05:21.689
of cause of your you're interested in
the ones who have yet to be sick,
74
00:05:21.730 --> 00:05:26.089
right, potential, and those are
the ones for which you have very
75
00:05:26.089 --> 00:05:28.209
little data, right, because if
you're not sick, you're obviously not going
76
00:05:28.250 --> 00:05:30.610
to go with the doctrine constantly be
measured for this man. So that's really
77
00:05:30.689 --> 00:05:38.000
the paradox. Where there is strong
data collect is amongst people for whom costs
78
00:05:38.040 --> 00:05:41.439
have already happened. Right. So
that as part of the de collect but
79
00:05:41.600 --> 00:05:45.439
that's not, in particular, where
your most interested if you're interested in reducing
80
00:05:45.480 --> 00:05:49.000
the cost of healthcare and making healthcare
more available to, in particular, seniors
81
00:05:49.000 --> 00:05:53.589
in the US. Okay, thanks, side description, and that's also as
82
00:05:53.629 --> 00:05:56.949
the couple to talk about, but
on experience plaining. Now, what is
83
00:05:56.990 --> 00:06:01.870
the situation in which the risks of
models are calculated based on individual outcomes and
84
00:06:02.069 --> 00:06:05.660
not aggregate outcomes? It's an excellent
question. So if you're dealing in a
85
00:06:05.779 --> 00:06:13.100
situation where the results of your data
science activity can dramatically impact an individuated person,
86
00:06:13.259 --> 00:06:16.019
right, you're not infecting people,
you're impacting persons. Right. So,
87
00:06:16.459 --> 00:06:23.529
for example, if my healthcare model
says that Mrs Jones should not yet
88
00:06:23.889 --> 00:06:28.970
like a particular surgery or should not
get a particular intervention, the potential downside
89
00:06:29.089 --> 00:06:30.529
from that could really be severe.
So, of the machine learning model,
90
00:06:30.529 --> 00:06:33.759
it's only rating on sort of the
minimum means squared error, I. which
91
00:06:33.759 --> 00:06:40.759
is sort of an average effect.
Right, then you've missed the individuated risk.
92
00:06:41.000 --> 00:06:44.560
That can happen, not just in
medical but in military applications or financial
93
00:06:44.639 --> 00:06:48.990
applications. If somebody doesn't get alone
or has some severe financial downside from the
94
00:06:49.029 --> 00:06:51.230
output. If your mode you can
go around saying Oh, well, on
95
00:06:51.310 --> 00:06:58.990
average worked right, because the severe
outcomes aren't captured sort of in the metric
96
00:06:59.069 --> 00:07:02.819
that's being used to to collect and
qualify the model. Right. So it's
97
00:07:02.819 --> 00:07:08.019
very seldom, and I'm actually never
seen it in practice where the out the
98
00:07:08.220 --> 00:07:13.579
mathematical objective of a machine running model
is sufficient to satisfy the organization seeds.
99
00:07:13.660 --> 00:07:17.129
I've actually never seen this. It's
always that the value that is a combination
100
00:07:17.370 --> 00:07:23.009
of the objectives of the machine learning
model but as their influenced by the other
101
00:07:23.129 --> 00:07:25.449
needs of the organization. So,
for example, we have a lot of
102
00:07:25.529 --> 00:07:29.250
privacy and security needs. Right,
no one's going to care about the outcome
103
00:07:29.250 --> 00:07:31.360
of this particular about how great it
is, how low the minimum means squared
104
00:07:31.399 --> 00:07:36.199
error is, if the other needs
of the organization unfolded into into how value
105
00:07:36.319 --> 00:07:41.439
is is, you know, recognized
by the model that we're going in a
106
00:07:41.519 --> 00:07:45.470
data and a little issues. I
guess what is the biggest challenge that you
107
00:07:45.589 --> 00:07:49.110
face and your industry? So I
think anybody dealing with sensitive data. I
108
00:07:49.230 --> 00:07:53.910
think security and privacy. As you
know, with all the constant news reports
109
00:07:53.949 --> 00:07:58.620
of everybody you know being held hostage
for data, being hijacked, etcetera.
110
00:07:58.740 --> 00:08:03.459
So anybody who deals with any kind
of secure information, in particular health conformation,
111
00:08:03.699 --> 00:08:09.579
big challenges for Security and privacy.
So maintaining trusting relationships with the other
112
00:08:09.779 --> 00:08:11.300
collaborators that you have not just that
you say, Oh, you know,
113
00:08:11.379 --> 00:08:15.730
we're secure and we're good. You
have to be in a situation where you
114
00:08:15.769 --> 00:08:20.490
have policies and procedures in place that
can demonstrate artifacts that that effect, that
115
00:08:20.649 --> 00:08:24.529
can head off problems before they happen. So I think security and privacy are
116
00:08:24.930 --> 00:08:28.040
big issues and are going to become
bigger issues as more data becomes a or
117
00:08:28.120 --> 00:08:33.440
the value of data in so far
as it recognizes issues of privacy and security
118
00:08:33.720 --> 00:08:37.159
become more common. And then the
other thing that we have is data collection
119
00:08:37.279 --> 00:08:41.399
is not for your cheap, in
particular in this area. So the ask
120
00:08:41.789 --> 00:08:46.429
for collecting data that directly influence as
a stated goal is something yet to be
121
00:08:46.590 --> 00:08:52.750
cognizant of. So collecting data actually
that's that drives towards a specifical isn't so
122
00:08:52.870 --> 00:08:56.340
easy when you're when you're dealing with
with healthcare dat so and the world I
123
00:08:56.419 --> 00:09:00.980
work in, those two most dominant. I think the security and privacy one
124
00:09:01.100 --> 00:09:05.019
is going to become a bigger deal
when you're dealing with other other kinds of
125
00:09:05.100 --> 00:09:07.820
industries. I just think the healthcare
one is sort of, you know,
126
00:09:07.299 --> 00:09:11.970
healthcare of financial military. I mean
they're already up into that, but I
127
00:09:13.009 --> 00:09:16.690
think that's gonna become a bigger deal
in future. Security, privacy and collecting
128
00:09:16.730 --> 00:09:20.169
data. Okay, so we get
the data standard. Are Building a community
129
00:09:20.210 --> 00:09:24.360
where data enthusiasts to collaborate things,
to connect. And so what can we
130
00:09:24.440 --> 00:09:30.000
do for you? So, Um, there's a great overview of the research
131
00:09:30.159 --> 00:09:33.360
that we do here at the institute
on the West Health Dot org site.
132
00:09:33.440 --> 00:09:35.840
So that's a lot. That's a
great place to keep up sort of on
133
00:09:35.879 --> 00:09:37.590
the overall. There's a lot more
to it actually, I was able to
134
00:09:37.629 --> 00:09:41.950
get into we also have a Github
site where we have where we support a
135
00:09:41.990 --> 00:09:45.230
number of open source products, in
particular visualization products. So you can go
136
00:09:45.389 --> 00:09:50.110
to gethubcom slash the West health you
can see some of the open collaborations we
137
00:09:50.230 --> 00:09:54.299
have regarding visual relation tools in particular, and other things that we that we
138
00:09:54.379 --> 00:09:58.100
do for the community. And then, if you're interested, there's also my
139
00:09:58.179 --> 00:10:03.259
own get up side. Get UPCOM
and Pinko has all the Jupiter notebooks corresponding
140
00:10:03.259 --> 00:10:05.419
to the books I've written if you
want to connect. So there's plenty of
141
00:10:05.500 --> 00:10:09.649
ways you can collaborate with us on
a technical level and research level. So
142
00:10:09.210 --> 00:10:13.570
we're happy to talk to anybody's interested. Okay, thanks for the staying the
143
00:10:13.610 --> 00:10:18.210
websites and then also work in viewers. Find you at that same website you
144
00:10:18.289 --> 00:10:22.919
just listed. I get home or
Christian, so it's a githubcom and then
145
00:10:22.440 --> 00:10:28.799
the last name it's you and pausing
Peter I and is never g CEO,
146
00:10:28.519 --> 00:10:31.720
and you'll find all the other material
I have on there. Involved in a
147
00:10:31.879 --> 00:10:35.190
number of other open source projects as
well. So yeah, we're happy to
148
00:10:35.350 --> 00:10:39.470
collaborate in where we have were always
we have papers we put out of the
149
00:10:39.549 --> 00:10:45.110
various python conferences. So we're big
python shop here. So well, thank
150
00:10:45.149 --> 00:10:48.470
you and thanks for, yeah,
beings allowing to be kind of the community.
151
00:10:48.950 --> 00:10:54.100
So you can find us at the
data standard www dot data standard ioh
152
00:10:54.460 --> 00:10:58.779
this episode was also sponsored by Pandio
and they're innovating the text base and you
153
00:10:58.820 --> 00:11:01.899
can find them at CANOCOM. But
back to is a thank you so much.
154
00:11:03.220 --> 00:11:05.690
You're insided for being here. Today, oh, thanks, thanks for
155
00:11:05.690 --> 00:11:07.169
your opportunity. I really pushing
