calculations kudos einsteinascrs16.expoplanner.com/handouts_ascrs/001048_37120624...13& general...
TRANSCRIPT
3/2/16&
1&
Data-Driven IOL Calculations Gerald Clarke, M.D. Warren Hill, M.D.
Financial Disclosures Gerald Clarke, M.D. Owns Copyright, CMO - FullMonteData, LLC Consultant, Appasamy Associates Founder, CEO, ReaLens, Inc.
Warren Hill, M.D. The services of MathWorks were provided through a research grant from Haag-Streit,
Switzerland. The author’s current industry relationships are with: Alcon Laboratories Consultant, Speaker
Clarity Medical Systems Consultant, Stockholder
Haag-Streit Consultant, Speaker, Research Oculus Consultant, Research
Kudos
Surgeons and Staff at: Florida Eye - Val Zudans, M.D. Chippewa Valley Eye - Tom Harvey, M.D. Montefiore Eye Center, Jimmy Lee, M.D., Rebecca
Weiss, Philip Kurochkin, Matthew Nicholas Eye Doctors of Washington, Paul Kang, M.D. Adam Kapelner, Ph.D., Assistant Professor Mathematics,
Queens College, NYU
The Great Ones: Warren Hill, Jack Holladay, Ken Hoffer, Don Sanders, Retzlaff, Manus Kraff, Wolfgang Haigis, Thomas Olsen, Graham Barrett, and Many More..
Einstein:
As far as the laws of mathematics refer to reality, they are not certain;
and as far as they are certain, they do not refer to reality.
3/2/16&
2&
All Models are Wrong!!
Main Question: How Wrong?
Models are Representations of Reality
Current IOL Calc Models: SRK-T, HofferQ, Holladay 1&2, Haigis, Barrett, Olsen
Current Models
Formula calculations Published – SRK-T, Holladay 1, HofferQ, Haigis
UnPublished – Holladay 2, Barrett, Olsen
Data Driven Computed – Not Calculated
Store Large Data Sets & Algorithms On Machine (Biometers – Haag Streit)
In Cloud (FullMonte IOL)
Future of Our Numeric Specialty
The Problem: Precision
Colt 0.45 Walther P.22
The Problem: Precision
Colt 0.45 Walther P.22
The Standard Deviation in all Current Formulas =
0.45 – I want 0.22
3/2/16&
3&
The Other Problem: Outliers
5% > 1.0 Diopters Pred Error
0.1% > 1.0 Diopter Prediciton Error
Current Models
Formula calculations ! a Single Number
Data Driven " A qualified Number or Probability Distribution
Future of Our Numeric Specialty
Most Errors outside Sweet Zone
Most Errors outside Sweet Zone
Short $ LOng ACD => Errors
3/2/16&
4&
What Do the Best Models Do?
Predict Well on UnSeen (out of sample) Data
Robust (Don’t Break)
Handle Multi-variable data
Handle missing data
Indicate Quality or Probability of Prediction (So Your Brain Machine Can Decide)
Predict Unseen Data
Single Most important Function of a Model
How Well Does the Model work on out of Sample (never been seen) data?
Predict Unseen Data Predict Unseen Data
3/2/16&
5&
Predict Unseen Data Predict Unseen Data
Predict Unseen Data Predict Unseen Data
3/2/16&
6&
Predict Unseen Data
SRKT HOLL HAIGIS Grnn MCMC
Stand Dev
0.45 0.44 0.60 0.40 0.25 – 0.35
Predict with MultiVariate Data
Astigmatism
Corneal Shape or Aberrations, Anterior and Posterior Cornea, all Meridians
Pupil Size
WTW, internal STS
Papa John:
Better Ingredients =
Better Pizza
MultiVariate Data
3/2/16&
7&
MultiVariate Data MultiVariate Data
MultiVariate Data MultiVariate Data WtWtW/
STS Ubm STS
3/2/16&
8&
MultiVariate Data
Ubm STSd
MultiVariate Data
MultiVariate Data MultiVariate Data
3/2/16&
9&
Ubm Data Keratomety Data
Q- Shapefactor (+ Pupil Size) – Hoffer/Savini
Anterior/Posterior Cornea Multiple Zones
Zernike Aberrations
Good Models Handle Many Variables
Add New Variables as needed
Handles Complexity
Factors (Post Lasik, DSAEK) can bifurcate predictions
Good Model: Robust to Outliers
All Formulas work reasonably well in central average area
Predictions Break Down for Long/short eyes/ Extreme Ks
Machine Learning Models tend to produce predictions that revert to mean
3/2/16&
10&
Good Model: Robust to Missing Data
Important Feature – Not all data can be collected with current biometry/keratometry (Lens Thick)
MCMC models are re-sizable, dropping or adding variables as needed. Do not Break with holes in data sets. Many Neural Nets can have same features
Robust to Missing Data
axial_len
gth& acd&
wt
w& lensthick&
re=nalthi
ckness&
q_shapefact
or&
pup
il& cct& sts& sts_d& rc& rc1&
rc2_2m
m&
rc1_2m
m&
rc1_4m
m&
rc2_4m
m& rp& rp1& rp2&
rpsteep
axis&
23.78& 3.53&
11
.6
8& 4.51&NULL& H0.21& 4& 550& 12& 4&
7.53432
3027&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
22.37& 3.61&
11
.6
8& 4.81&NULL& H0.25& 4& 550& 12& 4&
7.15421
3037&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
23.99& 3.37&
11
.6
8& 4.41&NULL& H0.06& 4& 550& 14.65& 4.04&
8.00426
8943&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
23.93& 3.27&
11
.6
8& 4.57&NULL& H0.36& 4& 550& 14.75& 4.17&
8.09449
5743&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
24.21& 2.91&
11
.6
8& 4.84&NULL& H0.52& 4& 550& 13.36& 3.47&
7.85980
4378&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
23.08& 2.81&
11
.6
8& 5.08&NULL& H0.09& 4& 550& 9.87& 3.7&
7.85705
9714&NULL& NULL& NULL& 0& 0& 0&NULL& NULL& NULL&
23.11& 2.79&
11
.6
8& 5.15&NULL& H0.78& 4& 550& 9.94& 3.34&
7.79625
7796&NULL&
23.27& 3.12&
11
.6
8& 4.83&NULL& H0.32& 4& 550& 10.72& 3.63&
7.63920
3259&NULL&
Good Models -> How good is my prediction?
Quality of Prediction
Assumptions underlying Predictions SRKT –assumes non-imaginary ACD
Haigis – assumes ELP Linear to ACD + AXL
MCMC models assume data at least T-distributed (Fat Tails)
How good is my prediction?
3/2/16&
11&
How good is my prediction?
85%
Your Learning Machine (Intellect)
Can decide to use the model's predictions or not
Algorithms of Machine Learning
Algorithms of Machine Learning
Big Data (Google-plex)
Machine Learning Competitions
Amazon Machine Learning Service
3/2/16&
12&
Algorithms of Machine Learning
Continuous Predictors: (vs Classifiers)
Linear and Non-Linear Regression
GRNN $ Neural Nets
Support Vector Machines
Genetic Algorithms
MCMC predictions
Linear and Non-Linear regression
Non-Linear regression Neural Nets
Warren Will Discuss
3/2/16&
13&
General Regression Neural Nets
Statistical model based on Gaussian Kernels (Normals)
Support Vector Machines
Markov Chain Monte Carlo MCMC Techniques
Multi-dimensional integrals (NERD STUFF) = Random Walks
When an MCMC method is used for approximating a multi-dimensional integral, an ensemble of "walkers" move around randomly. At each point where a walker steps, the integrand value at that point is counted towards the integral. The walker then may make a number of tentative steps around the area, looking for a place with a reasonably high contribution to the integral to move into next.
Theorem: MCMC can converge to a final probability distribution
Bayesian Prediction
Bayes Principles
Predictions with Assumptions
3/2/16&
14&
Bayesian Inference
Prior
Evidence
Most Likely Outcome
Tree Models
Traversing Trees a billion times
Practical Matters Does It Work?
Clinical Studies
Continually refine Results
Track Outcomes
3/2/16&
15&
Clinical Results 2013 Dr. Val Zudans
Prospective Trial - Used MCMC Guided IOL Selection (FMG) vs SRKT Chose IOL +/-0.25 diopters in Direction of MCMC Predicted Formula
(good Bayesian Decision) N=93
Results – Reduced StdDev from 0.44 (SRKT) to 0.36 (FMG)
Significant – at p-value = 0.05478 FMG better predictor 53 times – versus 21 for SRKT, 18 equal
90% of Eyes within 0.50 diopters of intended refraction 61% of Eyes within 0.25 diopters of intended refraction
Clinical Results 2015 Drs. Clarke, Lee, Kapelner
Lens Model N Patients Mean Average Error
SofTecHD 1638 0.326
TecnisZ9002 1147 0.287 SN60WF 561 0.587
TecnisZMB00 470 0.251 ToricSN6AT_2-9 454 0.365 Tecnis1ZCB00 383 0.503 LI61AOSofPort 201 0.543 AcrysofReSTORSN6AD 172 0.307 ATm50SEcrystalens 172 0.334 SofTecHD_O 155 0.334 AkreosAOMI60 120 0.333
other 187 0.563
The IOL Power Calculation Process…
1. Get Patient History ?Refractive Surgery ?Dry Eyes, Corneal Disease ?Contact Lens wear ?Gotchas?
2. Send Data from LenStar or IOLMaster to Cloud (fullmonteiol.com) 3. Use The Probability Estimates to Pick Your IOL 4. Collect Postop Data, to Refine A Constant, C- Constant, Haigis Vectors
59
Back to the Future
Predict Astigmatism – Use total Corneal Astigmatism – Incorporate Ray-Tracing and aberrometry
Femto AKs – Better Nomograms
Groups (Factors)
Post Lasik/RK
DMEK, DSAEK
Lasik – all Flavors
3/2/16&
16&
Thanks for your attention
www.FullMonteiol.com
Smartest Guy in the Room