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UnPublished – Holladay 2, Barrett, Olsen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

Gerry@fullmonteiol.com

www.FullMonteiol.com

Smartest Guy in the Room