Age Stratified Risk Prediction of
Invasive versus In-situ Breast
Cancer: A Logistic Regression
Model
Mehmet Ayvaci1,2
Oguzhan Alagoz1,Jagpreet Chhatwal3,
Mary Lindstrom4,Houssam Nassif5,Elizabeth S Burnside2
1Industrial
and Systems Engineering, UW-Madison
2Radiology, UW-Madison
3Merck Research Labaratories
4Biostatistics
5Computer Science, UW-Madison
Breast Anatomy
Breast profile:
– A ducts
– B lobules
– C dilated section of duct to hold
milk
– D nipple
– E fat
– F pectoralis major muscle
– G chest wall/rib cage
• Enlargement:
– A normal duct cells
– B basement membrane
– C lumen (center of duct)
www.breastcancer.org
Progression of Breast Cancer
Atypical ductal
hyperplasia
Normal duct
Typical ductal
hyperplasia
Invasive ductal carcinoma
Ductal carcinoma in situ
(DCIS)
Age Stratified Risk Prediction of
Invasive vs In-situ Breast Cancer:
A Logistic Regression Model
Age Stratification for
Invasive vs In-situ Breast Cancer
• Primary modality of screening or diagnosis:
Mammography
• Performs differently in different age groups
• Sensitivity: Age
– <40  54%
– 40-49  77%
– 50-65  78%
– >65  81%
• Sensitivity: Breast Density
– 68% vs. 85%
– Younger vs. older
Age Stratification for
Invasive vs In-situ Breast Cancer
• Primary modality of detecting type of breast cancer:
Biopsy
• Incidence of DCIS has increased since adoption
of mammography
• DCIS has favorable prognosis: will
often not cause mortality for years
• PPV of biopsy 20%
Age Stratification for
Invasive vs In-situ Breast Cancer
• Invasive vs. DCIS distinction important because:
• Requires different treatment
• Life expectancy difference in older and younger
women
– Over diagnosis which does not correspond to
reduced mortality
– Breast cancer less aggressive in older women
– Invasive procedures more risky in older women
– Resources could be better spent on more
serious co-morbidities
Purpose and Methods
• Develop a risk prediction
model for prospective
differentiation of DCIS
versus invasive breast
cancer
• Measure and compare
model performance for
different age groups
LOGISTIC REGRESSION
1
N
1 e
 (    i X i )
ROC Curves
i 1
Purpose and Methods Contd.
Risk
Assessment
Tools
Clinical
Implications
ROC & PR Curves,
Statistical Testing
Markov
Decision
Processes
Clinical
Implications
Structure of Data Used
NMD
National Mammography
Database Format
Structured
Demographic
Factors
Mammographic
Descriptors
BIRADS descriptors
Radiologists’ Overall
Assessment of the
Mammogram with Some
Repeat to the Structured Part
Free Text
Turned into Structured format
using Natural Language
Processing
Methods: Processing Free Text
• Information retrieval from free text given a
standardized lexicon
• Parse sentences to detect BIRADS descriptors using
Natural Language Processing in PERL
• Test on a set of 100 which is manually populated
– 97.7% Precision
– 95.5% Recall
Data in Detail
Structured








DEMOGRAPHIC

Age

Family History
Personal History
Prior Surgery
Palpable Lump
Interpreting Radiologist
Breast Density
Indication for Exam if Diagnostic

Associated Findings
Free Text
Calcification Distribution


Calcification Morphology
==
Features
Extracted
Using NLP
Mass Margins
Mass Shape
Mass Density
Mass Size
Special Cases
Biopsy






MAMMOGRAPHIC
BI-RADS Assessment
Principle Abnormal Finding
o Architectural Distortion
o Calcifications
o Asymmetry (one view)
o Focal asymmetry (two views)
o Developing asymmetry
o Mass
o Single Dilated Duct
o Both Calcifications and Something Else
Skin Retraction, Nipple Retraction, Skin Thickening,
Trabecular Thickening, Skin Lesion, Axillary Adenopathy,
Architectural Distortion
Clustered, Linear, Regional, Scattered, Segmental, Diffuse
Pleomorphic , Dystrophic, Lucent, Punctate, Vascular,
Eggshell, Dermal, Fine-Linear, Round, Popcorn, Milk of
Calcium, Rod Like, Suture, Amorphous,
Circumscribed, Obscured, Defined, Microlobulated,
Spiculated
Irregular, Lobular, Oval, Round
Fat Containing, Low Density, Equal Density, High Density
<10mm, >10mm and <20mm, >20mm and<50mm,
>50mm
Intra-mammary Lymph Node, Tubular Density, Assymetric
Breast Tissue, Focal Asymmetric Density
Insitu, Invasive
Summary of Data
• 1475 Diagnostic Mammograms  1378 Patients
– 1298 patients with single mammogram
– 81 patients with 2 mammograms
– 5 patients with three mammograms
• 1063 cases invasive vs. 412 DCIS
• Age range  27 to 97 with
– Mean 59.7 and standard deviation 13.4
Methods: Performing Logistic
Regression
• Regress with a dichotomous outcome, where the patient is known to have
malignant condition, i.e.
– Invasive or
– DCIS
P(Invasive|Demographic Factors, Mammographic Descriptors)
• Stratified data into 3 groups
– Overall Model LR 1475 records
– Age Less Than 50  LRyoung 374 records
– Age Greater Than 65  LROld 533 records
• Used stepwise regression to find the appropriate models. Possibility of
interactions were investigated
Methods: Validation Technique


n fold cross-validation
Leave-one-out
Fold 
Data set
1
2
3
…
n

…
Test fold
Training fold
Merge tested folds for performance analysis
Fold 

1
2
3
…
n
Methods: Measuring Performance
Sensitivity vs. 1-Specificity at all
thresholds
– Sensitivity: True Positive Rate
– Specificity: True Negative Rate
– Thresholds: Probability above
which call “Invasive”
– AUC: Area Under the Curve
Patient with
Disease
Patients
without disease
Test +
a
b
Test -
c
d
Sensitivity=a/(a+c)
Specificity = d/(b+d)
Results: LR
Age Based Error Rates
35.00%
30.00%
ROC
32.58%
100.00%
30.00%
90.00%
25.00%
20.00%
22.73%
22.35%
19.70%
17.64%
17.61%
15.58%
15.00%
12.72%
Sensitivity (TPR)
80.00%
70.00%
60.00%
AUC 0.843
50.00%
40.00%
30.00%
20.00%
10.00%
10.00%
0.00%
0.00%
5.00%
0.00%
AgeLT50
Age5064
FPR
FNR
AgeGTE65
Total Error Rate
Row Labels
Subject Count
Actual Invasive Count
AgeLT50
Age5064
AgeGTE65
Overall
374
568
533
1475
264
398
401
1063
20.00%
40.00% 60.00%
1-Specifity (FPR)
80.00%
100.00%
• Overall model significant at pvalue<0.01
• Not enough power to justify inclusion
of interaction terms (Over-fitting)
• Acceptable ROC
• Decreasing trend in Error rates
Results: LRyoung vs. LRold
Risk Factor
Significant in
Both Models
Significant in
Only Young
Younger
Odds
P-Value
Palpable Lump
>1
0.01
>1
<0.001
Mass Shape
~>1
0.03
~>1
0.03
Principal Finding
>1
<0.001
>1
<0.001
Architectural
Distortion
>1
0.06
Varies
0.05
Varies
0.04
Calcification
Distribution
<1
0.01
Focal Asymmetry
>1
0.08
Prior Surgery
>1
0.13
Mass Size
Family History
Significant in
Only Old
Older
Odds
P-Value
Results: LRyoung vs. LRold
• Difference in AUC =
0.07
• Significant at p-value
= 0.045
ROC Older vs. Younger
100.00%
90.00%
80.00%
Sensitivity
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
0.00%
20.00%
40.00%
60.00%
1-Specifity
80.00%
Age LT 50 AUC:0.778
Age GTE 65 AUC:0.848
100.00%
Results: LRyoung vs. LRold
• Improvement is in
False Negatives
False Rate Over Threshold
100.00%
90.00%
80.00%
70.00%
Rate
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Threshold
FNR Age GTE65
FPR Age LT50
FNR Age LT50
FPR Age GTE65
0.8
0.9
1
In Summary
• Mammography is not perfect and performs better in older
women.
• There is a need for discriminating between invasive and DCIS
to better manage the breast disease in the context of age and
other comorbidities
• An age based risk prediction model for assessing performance
difference in discriminating invasive vs. DCIS is necessary
• Such a model would enable physicians to make more
informed decisions
• Demonstration of performance difference and varying risk
factors in different age cohorts justifies
Future Work
Risk
Assessment
Tools
ROC & PR Curves,
Statistical Testing
Markov
Decision
Processes
Clinical
Implications
Get in
Literature
Clinical
Implications
Using POMDPs to Determine the Optimal Mammography Screening Schedule From the Patient's Perspective Presenting Author:
Turgay Ayer,University of Wisconsin
Co-Author: Oguzhan Alagoz,Assistant Professor, University of Wisconsin-Madison
Questions?
THANK YOU!