multi-state model for the estimation of overdiagnosis in ... · screening detects a breast cancer...
TRANSCRIPT
Multi-state Model for the Estimation of Overdiagnosisin the Breast Cancer Screening Program
Wendy Yi-Ying Wu
Department of Radiation Sciences, Oncology
Screening detects a breast cancer that would not have presented clinically in the woman's lifetime in the absence of screening
Overdiagnosis
Screening
Prolong lifeTreatmentPictures from Duivenvoorden HM et al., (2018) PlosONE, 13(7):e0201370Tabar L et al., (2005). Breast Cancer, Thieme
Qantification is difficult!Estimates varies from 0-56% (EUROSCREEN)
4
Control group?
Evaluation method?
Lead time adjustment?
Useful method for service-screening? Definition?
Denominators?
Illustration
5
Can
cer
prog
ress
ion
(Siz
e) Subject 1
Became detectable
Symptomatic
Time
Subject 2
BC: Breast cancer, PCDP: Preclinical screen-detected phase, CP: Clinical phase, SD: Screen-detected case
PCDP
Free of BC
CP
Non-progressive PCDP
screening
SD
Sojourn time for subject 1
Multi-state model
6
S1 1S1-S 1-S
: transitionrate from state to state at time t, S: test sensitivityBC: breast cancer, PCDP: preclinical screen-detectable phase, CP: clinical phase
Observed state Y(t)
Normal finding(Observed state 1)
Screen-detected cases
(Observed state 2)
Clinically detected cases
(Observed state 3)
Free of BC(State 1)
Progressive PCDP(State 2)
Non-progressive PCDP(State 4)
CP(State 3)Invisible
latent state X(t)
Screening registry
• PID• Date of
invitation• Date of screening• Screening results• Follow-up
(recall) results
• PID• Date of
diagnosis• Biopsy • Pathological
result
Cancer registry
Observed states{Y(t)}
Likelihood function
Estimate parameters
Calculate the expected number of
detected non-progressive BC
Data sources and work process
• Panel data (longitudinal follow-up data)o State only observed at a finite number of times (interval
censoring)o Don’t know the state between times
Screening Data
• Left truncationo Only the asymptomatic subjects will be invited
• Time-inhomogeneous modelo The incidence rate depends on the age
• Measurement erroro Sensitivity ≠ 100%
• Informative censoring ??o The subjects with shorter time to clinical phase tend to be
clinically detected if the participation rate is low
Screening DataConditional probability
Piece-wise rate/ parametric
Hidden Markov model
EM algorithm??
11Conclusion: Both methods can provide a proper estimate
True Cumulative incidence method
Multi-state model
Randomized controlled trial
1: 1 ratio screened and control, 100% participation rate
Control group provided information to stabilize the model. (help to deal with the identifiability problem)
Control group in the service screening program?
• Never-attenders (who were invited but never participated in the screening program) can provide the similar information under certain conditions, i.e. similar incidence rate
o Group A: those who will never attend any round of screeningo Group B: those who might attend the screening but does not
participate in the current/previous rounds Shorter time to CP few rounds of invitationmore clinical
cases (observation process and disease process are not independent= informative censoring??)
Population
Screened cohort
Controls / never-
attenders
PNA
1-PNA
1st screening
attenders
Not attenders
Women Invited
PPART
1- PPART
Women Invited
SD IN
NP
2nd screening
…
Women without
BC
NPSD: screen-detected, IN: interval cancer, NP: non-participant
A naive method (removing all the never-attenders)
Given on PNA=0.1
Non-homogeneous model Homogeneous model
Future studiesMethod
• Develop a flexible multi-state model
• Age, period or cohort effect
• Informative censoring
Sweden
• Hormone replacement therapy
• Digital mammography
• Different evaluation methods
Collaboration
• Norway
• Finland
• …
THANKS FOR YOUR ATTENTION
18
• Acknowledgemento Umeå University
Håkan Jonsson Lennarth Nyström
o Stockholm-Gotland Regional Cancer Centre Sven Törnberg Klara Miriam Elfström Anna Stoltenberg
Contact information:[email protected]