-
Notifications
You must be signed in to change notification settings - Fork 8
/
discr.html
470 lines (354 loc) · 15.6 KB
/
discr.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
<!DOCTYPE html>
<html>
<head>
<title>Discriminability</title>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />
<link rel="stylesheet" href="fonts/quadon/quadon.css">
<link rel="stylesheet" href="fonts/gentona/gentona.css">
<link rel="stylesheet" href="slides_style_i.css">
<script type="text/javascript" src="assets/plotly/plotly-latest.min.js"></script>
</head>
<body>
<textarea id="source">
<!-- TODO add slide numbers & maybe slide name -->
### Eliminating Accidental Deviations in Human Connectomics
![:scale 40%](images/neurodata_blue.png)
Joshua T. Vogelstein | {[BME](https://www.bme.jhu.edu/),[CIS](http://cis.jhu.edu/), [ICM](https://icm.jhu.edu/), [KNDI](http://kavlijhu.org/)}@[JHU](https://www.jhu.edu/) | [neurodata](https://neurodata.io)
<br>
[jovo@jhu.edu](mailto:[email protected]) | <http://neurodata.io/talks> | [@neuro_data](https://twitter.com/neuro_data)
---
### Outline
- [Motivation](#defn)
- [Quantifying Discriminability](#statistics)
- [Properties of Discriminability](#props)
- [Real Data](#results)
- [Discussion](#disc)
---
name:defn
### Outline
- Motivation
- [Quantifying Discriminability](#statistics)
- [Properties of Discriminability](#props)
- [Real Data](#results)
- [Discussion](#disc)
---
### What is Reproducibility?
- .ye[Reproducibility]: ability to replicate, or reproduce, a conclusion
- serves as a "first-pass" check for scientific utility
- currently in a "reproducibility crisis"
---
### How do we address the Reproducibility Crisis?
- fix post hoc analyses (e.g., $p$-values)?
- fix measurements (e.g., measurement reproducibility)?
Proposal: design experiments to maximize .ye[inter-item discriminability], rather than simply checking reproducibility after conducting the experiment
---
name:statistics
### Outline
- [Motivation](#defn)
- Quantifying Discriminability
- [Properties of Discriminability](#props)
- [Real Data](#results)
- [Discussion](#disc)
---
### What do we want of our data?
If we measure a sample multiple times, then each measurement of that sample is closer to all the other measurements of that sample, as compared to any of the measurements of other samples.
![:scale 100%](images/discr/perfect_discrim.png)
Perfect discriminability
---
### What do we want of our data?
Imperfect discriminability
![:scale 100%](images/discr/imperfect_discr.png)
---
### What do we want of our statistic?
Discriminability is the probability of a measurement from a different item being closer than a measurement from the same item.
![:scale 100%](images/discr/imperfect_discr.png)
<!-- --- -->
<!-- ### Statistics for Quantifying Reproducibility -->
<!-- - most approaches tend to focus on the concepts of between-individual and within-individual summaries -->
<!-- -- -->
<!-- - parametric approaches -->
<!-- - reproducibility is achieved if most of the variability is between-individuals, rather than within-individuals -->
<!-- - e.g. Intraclass Correlation Coefficient (.ye[ICC]) and Image Intraclass Correlation Coefficient (.ye[I2C2]) -->
<!-- -- -->
<!-- - non-parametric approaches
- reproducibility is achieved if measurements from the same item are more similar to one another than measurements from different items
- e.g. Fingerprinting (.ye[Finger.]) and DISCO (.ye[Kernel]) -->
<!-- --- -->
<!-- ### Limitations of Existing Reproducibility Statistics -->
<!-- - Data: $N$ items in $d > 1$ dimensions measured $s \geq 2$ times each -->
<!-- | Statistic | Chief Limitation | -->
<!-- | :--- | :--- | -->
<!-- | ICC | univariate -->
<!-- | I2C2 | does not generalize outside Gaussian framework; estimation issues in high dimensions -->
<!-- | Finger. | overly greedy: only look at most similar measurement to the reference measurement -->
<!-- | DISCO | struggles when $s$ is small -->
<!-- --- -->
<!-- ### Discriminability Statistic: Data -->
<!-- - $n$ individuals measured $s = 2$ times each, for $N = n\cdot s$ total measurements -->
<!-- - generalizes to $s \geq 2$, but $s=2$ is simpler -->
<!-- <img src="images/discr/dummy_sim_sim.png"/> -->
---
### Discriminability Statistic: Step 1
- Compute $N \times N$ pairwise distance matrix between all measurements
<!-- - measurements are indexed first by individual identifier, and second by "measurement session" -->
![:scale 55%](images/discr/imperfect_discr.png)
![:scale 42%](images/discr/dummy_sim_dmtx.png)
<!-- <img src="images/discr/dummy_sim_dmtx.png"/> -->
---
### Discriminability Statistic: Step 2
- For each measurement, identify which measurements are from the same individual (<font color="green">green boxes</font>)
- let $\color{green}g$ be the total number of <font color="green">green boxes = 20</font>
<img src="images/discr/dummy_sim_dmtx_match.png"/>
---
### Discriminability Statistic: Step 3
- For each measurement, identify measurements from other individuals that are more similar than the measurement from the same individual (<font color="orange">orange boxes</font>)
- let $\color{orange}f$ be the total number of <font color="orange">orange boxes = 84</font>
<img src="images/discr/dummy_sim_dmtx_closer.png"/>
---
### Discriminability Statistic
- Discr = $1 - \frac{\color{orange}f}{N(N-1) - \color{green}g} = 1 - \frac{\color{orange}{84}}{20\cdot 19 - \color{green}{20}} \approx .77$
<img src="images/discr/dummy_sim_repr.png"/>
High discriminability: same-item measurements are more similar than across-item measurements
---
### Discriminability is Construct Valid
<img src="images/discr/dummy_sims.png" style="height: 350px"/>
<!-- - under the given construct (what the simulation is supposed to show), discriminability provides a sensible statistic -->
<!-- - other approaches do not -->
---
name:props
### Outline
- [Motivation](#defn)
- [Quantifying Discriminability](#statistics)
- Properties of Discriminability
- [Real Data](#results)
- [Discussion](#disc)
---
### Discriminability: unbiased and consistent
- population discriminability $D$ is a .ye[property of the distribution] of measurements
<!-- - A sequence of items $x_i^k$ from individuals $i=1,..., N$ measured at time $k=1,..., s$ -->
$D = \mathbb P(\delta(x_i^k, x_i^{k'}) < \delta(x_i^k, x_j^{k''}))$
<!-- - Probability that distances from measurements of the same individual are smaller than distances from measurements of different individuals -->
- Sample Discr. is an .ye[unbiased estimator] of $D$
- Sample Discr. converges to $D$ (.ye[asymptotically consistent])
---
### Simulation Setup
<!-- - Construct $3$ simple simulations where the data are describable in Gaussian (or non-Gaussian) framework -->
<!-- - level of "noise" in the simulation is varied -->
<!-- <center><img src="images/discr/sims_sim.png" style="height: 450px"/></center> -->
![:scale 70%](images/discr/sims_sim.png)
---
### Connecting Discriminability to Downstream Inference
<!-- - Fundamental question: does discriminability .ye[matter] for inference? -->
<!-- ##### Assumptions -->
<!-- Data follows Gaussian mixture model plus additive Gaussian noise -->
#### Theorem 1
$D$ provides a lower bound on the predictive accuracy of a subsequent classification task, and consequently
#### Corollary 2
A strategy with a higher $D$ provably provides a higher bound on predictive accuracy than a strategy with a lower $D$
---
### Discriminability and Accuracy
![:scale 70%](images/discr/sims_acc.png)
<!-- <center><img src="images/discr/sims_acc.png" style="height: 450px"/></center> -->
Discr. decreases proportionally with accuracy
---
### Are data discriminable?
<!-- - Fundamental question: are the data discriminable at all? -->
![:scale 70%](images/discr/sims_os.png)
---
### Is one dataset more discriminable than another?
![:scale 70%](images/discr/sims_ts.png)
---
name:results
### Outline
- [Motivation](#defn)
- [Quantifying Discriminability](#statistics)
- [Properties of Discriminability](#props)
- Real Data
- [Discussion](#disc)
---
### What data will we be using?
- CoRR metadataset
- $N>1,700$ individuals imaged across $26$ different datasets
- anatomical MRI and fMRI scans for each
- Individuals are measured at least twice
---
### Analysis Procedure
Process each measurement using $192$ different pipelines
1. Brain alignment (ANTs/FSL)
2. Frequency filtering (Y/N)
3. Scrubbing (Y/N)
4. Global Signal Regression (Y/N)
5. Parcellation (4 options)
6. Rescaling connectomes (Raw, Log, Pass-to-Rank)
$192 = 2 \times 2 \times 2 \times 2 \times 4 \times 3$
All options represent strategies experts consider useful
---
### Pipeline impacts discriminability
<center><img src="images/discr/fmri_cmp.png" style="height: 420px"/></center>
<!-- - Consequence: choosing how to pre-process your data matters -->
---
### Marginally most discriminabile options tend to be best global options
<center><img src="images/discr/individual_methods.png" style="height: 200px"/></center>
- Best pipeline marginally (FNNGCP) is second best pipeline overall, and not much worse (2-sample test, p=.14) than the best pipeline FNNNCP
- We may not need to always try every pre-processing strategy every time
---
### Selection via Discriminability improves inference
For each pre-processing strategy, for each dataset, compute:
1. Within-dataset Discr.
2. Demographic effects (sex and age) within the dataset via Distance Correlation (DCorr)
3. Within a single dataset, regress demographic effect on Discr.
Question: does a higher discriminability tend to yield larger effects for known biological signals?
---
### Selection via Discriminability improves inference
<img src="images/discr/dependence.png" style="height: 400px"/>
<!-- - Consequence: maximizing discriminability in general improves downstream inference -->
---
name:disc
### Outline
- [Problem Space](#prob)
- [Defining and Quantifying Discriminability](#statistics)
- [Properties of Discriminability](#props)
- [Real Data](#results)
- Discussion
---
### Contributions
1. Discriminability quantifies the contributions of systematic and accidental deviations
3. Provide theoretical motivation for discriminability in connection with predictive accuracy
2. Formalize tests for assessing and comparing discriminabilities within and between collection strategies
4. Illustrate the value of discriminability for neuroscience and genomics (not discussed) sdata
5. Code implementations in [python](https://github.com/neurodata/hyppo) and [R](https://github.com/neurodata/r-mgc)
<!-- ---
### Connections to Energy Statistics
- energy statistics provide statistics that are functions of distances between measurements, like discriminability
- both approaches make relatively few assumptions
- discriminability 2-sample test is a comparison of K-sample tests in the energy statistics literature
- energy statistics tend to perform well when we have more replicates (e.g., DISCO) -->
<!-- --- -->
<!-- ### Extending Discriminability
- use of distances allows for trivial modification for non-Euclidean measurements
- e.g., speech or text data may make sense with different distance functions
- measurements that consist of multiple measurements themselves
- if we only have one measurement that consists of many sub-measurements, can split the sub-measurements into multiple measurements
- e.g., taking a single fMRI and splitting into two fMRI with $\frac{t}{2}$ timepoints each -->
---
### Limitations
- experimental design is not "one-size-fits-all"
<!-- - future scientific questions will still need to consider the question of interest -->
<!-- - e.g., an analysis of task fMRI may not want to use a pre-processing pipeline with global signal regression, but for resting-state fMRI, this may not be an issue -->
- Discriminability is a pre-requisite for practical utility
- categorical covariates are meaningful but not discriminable
- fingerprints are discriminable but not typically biological useful
- These statistics are not immune to sample characteristics
- confounds such as age may inflate discriminability
---
### Acknowledgements
<div class="small-container">
<img src="faces/ebridge.jpg"/>
<div class="centered">Eric Bridgeford</div>
</div>
<div class="small-container">
<img src="faces/shangsi.jpg"/>
<div class="centered">Shangsi Wang</div>
</div>
<div class="small-container">
<img src="faces/zhi.jpeg"/>
<div class="centered">Zhi Yang</div>
</div>
<div class="small-container">
<img src="faces/zeyi.jpeg"/>
<div class="centered">Zeyi Wang</div>
</div>
<div class="small-container">
<img src="faces/ting.jpeg"/>
<div class="centered">Ting Xu</div>
</div>
<div class="small-container">
<img src="faces/cc.jpg"/>
<div class="centered">Cameron Craddock</div>
</div>
<div class="small-container">
<img src="faces/jayanta.jpg"/>
<div class="centered">Jayanta Dey</div>
</div>
<div class="small-container">
<img src="faces/gkiar.jpg"/>
<div class="centered">Greg Kiar</div>
</div>
<div class="small-container">
<img src="faces/wgr.jpg"/>
<div class="centered">William Gray-Roncal</div>
</div>
<div class="small-container">
<img src="faces/carlo.jpeg"/>
<div class="centered">Carlo Colantuoni</div>
</div>
<div class="small-container">
<img src="faces/cdouville.jpeg"/>
<div class="centered">Christopher Douville</div>
</div>
<div class="small-container">
<img src="faces/snoble.jpeg"/>
<div class="centered">Stephanie Noble</div>
</div>
<div class="small-container">
<img src="faces/cep.png"/>
<div class="centered">Carey Priebe</div>
</div>
<div class="small-container">
<img src="faces/bcaffo.jpg"/>
<div class="centered">Brian Caffo</div>
</div>
<div class="small-container">
<img src="faces/mm.jpg"/>
<div class="centered">Michael Milham</div>
</div>
<div class="small-container">
<img src="faces/xinian.jpg"/>
<div class="centered">Xinian Zuo</div>
</div>
<span style="font-size:200%; color:red;">♥, 🦁, 👪, 🌎, 🌌</span>
</div>
<img src="images/funding/nsf_fpo.png" STYLE="HEIGHT:95px;"/>
<img src="images/funding/nih_fpo.png" STYLE="HEIGHT:95px;"/>
<img src="images/funding/darpa_fpo.png" STYLE=" HEIGHT:95px;"/>
<img src="images/funding/iarpa_fpo.jpg" STYLE="HEIGHT:95px;"/>
<img src="images/funding/KAVLI.jpg" STYLE="HEIGHT:95px;"/>
<img src="images/funding/schmidt.jpg" STYLE="HEIGHT:95px;"/>
- [BioRxiv manuscript](https://www.biorxiv.org/content/10.1101/802629v6)
- Code implementations in [python](https://github.com/neurodata/hyppo) and [R](https://github.com/neurodata/r-mgc)
</textarea>
<!-- <script src="https://gnab.github.io/remark/downloads/remark-latest.min.js"></script> -->
<!-- <script src="remark-latest.min.js"></script> -->
<script src="remark-latest.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/KaTeX/0.5.1/katex.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/KaTeX/0.5.1/contrib/auto-render.min.js"></script>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/KaTeX/0.5.1/katex.min.css">
<script type="text/javascript">
var options = {};
var renderMath = function () {
renderMathInElement(document.body);
// or if you want to use $...$ for math,
renderMathInElement(document.body, {
delimiters: [ // mind the order of delimiters(!?)
{ left: "$$", right: "$$", display: true },
{ left: "$", right: "$", display: false },
{ left: "\\[", right: "\\]", display: true },
{ left: "\\(", right: "\\)", display: false },
]
});
}
remark.macros.scale = function (percentage) {
var url = this;
return '<img src="' + url + '" style="width: ' + percentage + '" />';
};
// var slideshow = remark.create({
// Set the slideshow display ratio
// Default: '4:3'
// Alternatives: '16:9', ...
// {
// ratio: '16:9',
// });
var slideshow = remark.create(options, renderMath);
</script>
</body>
</html>