Post on 30-Apr-2015
description
hupertan.stat@me.com
NomogramsWhy, When, What, How use?..
….but
The 2nd World Congress on
Controversies in Urology (CURy)
Lisbon, Portugal, February 5- 8, 2009
Vincent HUPERTAN, M.D., MR
Lyon University - E.R.I.C. Knowledge engineering
hupertan.stat@me.com
hupertan.stat@me.com
Nomogram ≠ predictive model(PM)is the graphical representation of
mathematical relationships or laws (Etymology: Greek nomos = law)
or a graphical calculating device, a two-dimensional diagram designed to allow the approximate graphical computation of a function.
Fah
ren
heit
vs.
Cels
ius
scale
hupertan.stat@me.com
• This slideshow is a visual support for interventions Dr. Hupertan as expert or trainer during training seminars , courses for medical students, conferences or congresses .
• This slideshow created by Dr. Hupertan , MD , is intended primarily for health professionals in training ( medical students , interns and clinical leaders ) or not (doctors,... ) .
• This slide contains links to other sites.
• Conflict of interest : "no declared conflicts of interest "
• Using Slideshow : this slideshow can be downloaded , used while mentioning the author.
hupertan.stat@me.com
Predictive nomogram★ A device that suppose two
elements:
1. equation of an
event probability
2. specific functional
representation in a
graphic form
hupertan.stat@me.com
Why?1. The necessity to improve the decision making process in
oncology1. Clinical heterogeneity of cancers2. Importance quantity / quality of life ratio3. Perfect treatment = utopia⇒ maximize cancer control/ minimize treatment morbidity
2. Lack of performance in prediction of the clinical judgment (CJ)1. The clinician (experts) out-perform prediction classifiers= too much
weight on their own judgment2. Human mental process prove difficulties to use numbers3. Emotional considerations: particular cases are more “weighted” ⇒ Accuracy of the prediction of the PM >> CJ
3. Paucity of RCT data implying a lack of the “evidence”*)=> we should use data to improve the medical decision making process, implying more actively patient in that
*)Evidence Based Medicine
hupertan.stat@me.com
… yes, BUT:
1. Maximize cancer control/ minimize treatment morbidity OK, BUT:Does exist nomogram able to predict in same time cancer control
and treatment morbidity?How to predict cancer control: survival? surrogate end points?What means treatment morbidity in a statistical point of view: QoL
score? Erection function IIEF?
2. Accuracy of the prediction of the PM >> CJ OK, BUT: Y=f(X1, X2, X3, ..,Xi)! Y=[ Y1, Y2, Y3, ..,Yn ]=f(X1, X2, X3, ..,Xm)! m inputs => n outputs (social, familial, sexual....)
Þ Imply more actively patient in the medical decision making process OK BUT: well informed patient = associate probability to each possible outcome ? Let himself on the new to compute the risk hazard? What probability he will choice you?
hupertan.stat@me.com
When?
To inform the patient about the outcome that MIGHT BE!
…the fact that predict the issue will change:diagnostic procedurestreatment choice (alternative treatments, adjuvant
treatment exists) or treatment modalities (extension of the lymph-nodes dissection)
follow-up
hupertan.stat@me.com
What “nomogram” to choose?(nomogram specifications)
1. Functional representation of the nomogram:Ergonomy, simplicity
2. Nomogram core(PM):Output:
relevant for the clinical practice;Data set used for the learning process:
Patients: geographic area, academic centersPredictors:
variability(inter rather,within rather),standardization
colinearity? significative features?exhaustivity or parsimony?
Quality of data set(?), noise (?), missing data (?)
hupertan.stat@me.com
2. Nomogram core(PM):• Modeling tools:
• machine learning: neuronal nets, machine vector, induction graph, bayesian
• statistic : regression, Cox model• symbolic learning, rules induction
Validation:internal:
learning set-test setbootstrap, jackknife
external:academic/non-academic centerspublication bias (negatives) «invalidating nomogram»
What “nomogram” to choose?(nomogram specifications)
hupertan.stat@me.com
• Simplest possible• Linked to an actionable question• Modeling: statistics, significativity of the features• Good performance in prediction:
• Accuracy (validation in similar sample data)• Calibration • Discrimination: Harell c index or AUC ∈(0.7-0.8*)
• Generalizability• Updating models using my own data (e.g. using
bayesian technics and bootstrap)• Estimation by confidence interval
What “nomogram” I use?
c index >0.8 : memorized data? over-learning?
hupertan.stat@me.com
How to use predictive nomograms
• is difficult to use it as well!• after validation in your data or in identical
sample• using à confidence interval, and if possible
built-in on your data• we should dispose official recommendation
(E.A.U., A.F.U.)• for patients to be informed BY doctors• permanent updated with new data:
• new patients• new features: genomic and biomolecular data 1
2
12
hupertan.stat@me.com
«no nomogram will ever take the place of good clinical judgement and the well-informed patients.»
Robert W. Ross, Philip W. KantoffPredicting Outcomes in Prostate Cancer: How Many More Nomograms Do
Se Nedd? J.CLIN.ONCOL, 25,2077:3563-3564
13
13
hupertan.stat@me.com
Thanks to:• Pr Laurent Boccon-Gibod, Bichat Hospital
• Pr Jean-Hugues Chauchat, Knowledge engineering Labs,
(PhD Thesis Director)
The presentation it has been inspired by papers• Michael W.Kattan
• Philip W. Kantoff
• Frank E.Harell
• Rodolfo Montironi
• Robert W. Ross
• Peter T. Scardino
• Ashutosh Tewari
• Blaz Zupan
and many others 14
14
hupertan.stat@me.com
Case Nr 150y old, caucasian, Website designer
Benign prostatic hyperplasia (BPH) with LUTS : AUA-SI = 8 (moderate)
Erectile dysfunction (ED), IIEF15 (Erectile Function-domain)=7 (severe)
PSA=8 ng/ml, DRE=T1c
Transrectal ultrasound-guided biopsy of the prostate:
prostate volume=30;
Gleason score= 4+5;
3 positives cores on 12.
The patient says: «Using internet I found that the probability to be healed 5 years latter as 87% in the case of the surgery, and only about 73% in the case of the external beam radiation therapy. I choose the surgery!»
15
15
hupertan.stat@me.com
Case nr 1
As urologist in a non academic center do you
operate him?
How to explain the difference?
What confidence around the estimate?
Progression Free Probability Radical
Prostatectomy meant «Healed»?
rising PSA after surgery= after radiation therapy
16
16
50y, Gleason score= 4+5, PSA= 8 ng/ml,T1c
hupertan.stat@me.com
Case Nr 265y old, caucasian, statistician
Hemochromatosis
Benign prostatic hyperplasia (BPH) with LUTS : AUA-SI = 20 (severe)
Erectile dysfunction (ED), IIEF15 (Erectile Function-domain)=26 (mild)
PSA=30 ng/ml, DRE=T1c
2 previously biopsy of prostate= negatives
Transrectal ultrasound-guided biopsy of the prostate:
prostate volume= 65 cc
Gleason score= 3+3;
6 positives cores on 12.
17
17
hupertan.stat@me.com
Case Nr 2As urologist in a non academic center you explain that in the case of
the prostatectomy the 5 years progression free probability is 93%, and only about 72% in the case of the external beam radiation therapy.
The patient (statistician) ask you: «But if YOU are the surgeon, what are the estimation of the same progression free probability?»
You have no Idea about it!
93% as the nomogram predict, because the nomogram has
been validated
93% ± 5% (α, risque of error)
around 93%
18
18