USING DATA MINING TECHNIQUES FOR DIAGNOSIS AND PROGNOSIS OF CANCER DISEASE
Poster_Visser2015
1. / Department of Industrial Engineering & Innovation Sciences
/ Information Systems Group
Patient-specific Management of Adrenal Incidentalomas
J.B.R. Visser - U. Kaymak - R.J. Almeida - J.J. Visser - M. van der Meijden
From Predictive towards
Prescriptive Analytics
Introduction
Adrenal incidentaloma (AI)
An adrenal mass (> 1 cm) detected on
imaging studies performed for indications
other than adrenal disease [1]. Prevalence
of AI on CT is 1-4.1% [2].
Management of adrenal incidentalomas
The challenge is to recognize and treat the small percentage of
clinically relevant incidentalomas that pose a significant risk.
Clinically relevant findings are malignant, cause hormonal
hyperfunction or grow significantly.
Research objectives
There is demand for a patient-specific management strategy
that offers multiple moments at which we want to decide
whether or not to continue AI work-up. Two objectives that
guide us from predictive towards prescriptive analytics are:
1. Develop a clinical prediction model to predict if an
adrenal incidentaloma is clinically relevant.
2. Derive a prescriptive model for patient-specific
management of patients with adrenal incidentalomas.
Prescriptive modeling
Integrate prediction models in Erasmus MC daily practice
using decision curve analysis and BPMN 2.0 (figure 2).
Conclusion and Future Work
This research resulted in a prescriptive model based on three
predictive models that predict if a patient will have a clinically
relevant outcome at two moments during adrenal inciden-
taloma work-up. Direction for future research are:
»» Implement additional decision moments by refining the
dataset with imaging variables from CT follow-up.
»» Perform external validation and enlarge dataset.
References
[1] Mansmann et al., (2004). The Clinically Inapparent Adrenal Mass: Update in Diagnosis and
Management. Endocrine Reviews
[2] Davenport et al., (2011). The Prevalence of Adrenal Incidentaloma in Routine Clinical Practice.
Endocrine
Results
Predictive models
Perform very good on the over-sampled test sets. Results
on the original validation sets are good but fluctuating
Prescriptive model
Patients that are not clinically relevant whose work-up was
rightly stopped based on the validation data:
»» M1: 15% of 96 patients saved from unnecessary work-up
»» M2: 25% of 16 patients saved from unnecessary work-up
All clinically relevant patients remained in the work-up.
Prescriptive Model
ErasmusMC
Radiology
Radiology Radiology
Imaging follow-up 1:
Non-contrast CT
Imaging follow-up:
CT
growth or
malignant
End AI
work-up
3-6 months
12 months
3 CT in total?
growth or
malignant
Perform CT exam
(inspect images
immediately)
AI found?
Assess patient
situation
Run Prediction
Model M1
Start AI
work-up?
End AI
work-up
End AI
work-up
Referring
Department
Referring Department Referring Department
Request CT
abdomen/thorax
Surgery
Surgery Surgery
Adrenal vein
sampling
Adrenalectomy
End AI
work-up
Endocrinology
Endocrinology Endocrinology
Endocrinologist visit 1:
Evaluate results
Patient history and
physical examination
Aldosterone
needed? Endocrinologist visit 2:
Evaluate results
Hormonal
hyperfunction?
Age > 40 and
hypertension?
Size > 4-6 cm?
Run Prediction
Model M2
Continue
work-up?
End AI
work-up
Laboratory
Laboratory Laboratory
Perform lab tests:
CS, FEO
Perform lab test:
PA
1 day 1 day
no
yes
yes
no
yes
yes
no
yes
no
no
no
yes
no
no
yes
yes
Figure 2: Prescriptive model for patient-specific management of adrenal incidentalomas
Figure 1: Experimental design for prediction models M1
Methods
Patient identification & data collection
Apply text mining to all radiology reports from 2010 till 2012
resulting in 643 patients. Patient data, lab test data, medica-
tion data and work-up data are collected. Of these adrenal
incidentalomas 5.3% are clinically relevant.
Predictive modeling
Construct three prediction models at decision moments:
»» M1: At the moment of finding (figure 1)
»» M2: After biochemical screening
The experimental design is identical for both moments.