Longitudinal breast density as a marker of breast cancer risk

Transcription

1 Longitudinal breast density as a marker of breast cancer risk C. Armero (1), M. Rué (2), A. Forte (1), C. Forné (2), H. Perpiñán (1), M. Baré (3), and G. Gómez (4) (1) BIOstatnet and Universitat de València, (2) BIOstatnet and Universitat de Lleida, (3) Epidemiology and Assessment Unit UDIAT-Diagnostic Centre, Corporació Sanitària Parc Taulí, Sabadell, (4) BIOstatnet and Universitat Politècnica de Catalunya Joint Modeling and Beyond Universiteit Hasselt, April 14-15, 2016

2 Outline Background Scientific objective and study population Statistical methods Results Estimation Prediction Conclusions 2 / 26

3 Sección 1 Background 3 / 26

4 Background Breast density (BD) is a characteristic of the breast tissue that is reflected in mammograms. For most women, breast density decreases with age. The ordinal BI-RADS scale (American College of Radiology, 2013) is a commonly tool used to measure breast density. It categorizes breast density in four groups a (1): almost entirely fatty (low density), b (2): scattered fibroglandular densities (medium density), c (3): heterogeneously dense (high density), and d (4): extremely dense (very high density) Several studies have shown that high breast density is associated with an increased breast cancer (BC) risk. Some research (Huo et al., 2014) concluded that longitudinal measurement of mammographic density might be used to personalize breast cancer screening (Biomarker). 4 / 26

5 Sección 2 Scientific objective and material 5 / 26

6 Objective and material Objective: To study the association structure and intensity between mammographic longitudinal measures of breast density and breast cancer risk. Study population: 13,685 women attending a population-based screening program in the BC early-detection program in the Vallès Occidental East area in Catalonia (Spain), between October 1995 and June Women were followed for vital status or possible diagnosis of BC until December, 2013 (Baré et al., 2003, 2006, 2008). At study entry years old. No history of BC. Invited for a biennial mammographic exam. 6 / 26

7 Study variables and scales: Breast density and breast cancer data Data from the 13,685 women attending our population-based screening program. 81,621 mammographic measures of breast density with the subsequent identification and age of the woman. Survival time: Time from the initial age at entrance to the screening program until BC-diagnosis (Administrative censoring). Time scale: Age (50-69 years). Time zero: 50 years. Women who entered the programme after this age produced left truncated data. Baseline covariates: First-degree relative with BC: Yes-No. Prior breast procedures (breast biopsy, fine needle aspiration, cyst aspiration, breast reconstruction, lumpectomy and surgical treatment): Yes-No. 7 / 26

8 Data description women and mammograms Count 1000 BI RADS Number of screenings per women Years Prior breast procedures: 7.1 % (no diagnosis BC) and 13.2 % (diagnosis BC) First-degree relative with BC: 5.2 % (no diagnosis BC) and 9.8 % (diagnosis BC) 431 women diagnosed with BC (3.15 %) 8 / 26

9 Sección 3 Statistical methods 9 / 26

10 Statistical methodology Elements: Longitudinal data: Breast density data Survival data: Time from the initial age of the screening program to BC detection. Left truncation and administrative censoring. Objective: We want to construct a survival model to assess the risk of a diagnosis of breast cancer by means of baseline information and breast density temporal information. Put a joint model in your problem 10 / 26

11 Objective and material and methods Joint models Different approaches: conditional models, shared-parameter models, random-effects models, and joint latent class models. Shared-parameter models: 1 Longitudinal model for the longitudinal data y. 2 Survival model for the survival data s. 3 Connection between both models by means of a common vector of random-effects, b, which endows them with conditional independence. Frequentist f (y, s, b) = f (y, s b) f (b) Bayesian = f (y b) f (s b) f (b) f (y, s, b, θ, φ) = f (y, s b, θ, φ) f (b, θ, φ) = f (y b, θ, φ) f (s b, θ, φ) f (b θ, φ) π(θ, φ) = f (y b, θ) f (s b, θ) f (b φ) π(θ, φ) where θ and φ are the vector of parameters and hyperparameters, respectively. 11 / 26

12 The joint model, I Longitudinal submodel: A cumulative logit model for the longitudinal ordinal measures of breast density based on the idea of a continuous latent variable (Lunn et al., 2001; Luo et al., 2014). y ij breast density in the BI-RADS scale, y ij {1, 2, 3, 4}, of woman i, i = 1,..., n, at time t ij (age 50 + t ij ), j = 1,..., n i. We assume an underlying continuous latent variable yij that determines the breast density BI-RADS category of woman i at time t ij. Relationship between y ij and y ij y ij = k y ij (γ k 1, γ k ], y ij Logistic(m ij, s = 1) logit P(y ij > k) = m ij γ k, m it = β 0 + b i0 + (β 1 + b i1 )t, with = γ 0 < γ 1 < γ 2 = 0 < γ 3 < γ 4 = unknown cutpoints, β 0, β 1 unknown parameters and (b 0i, b 1i subject specific random effects). 12 / 26

13 The joint model, II Survival submodel: A left-truncated Cox proportional hazard model for the time-to-bc detection that incorporates information from the longitudinal process. T i, i = 1,..., n observed BC detection time for the i-th woman, T i = min(ti, C i), Ti is the true failure time and C i the right-censoring time. The event indicator, δ i = I(Ti C i). Hazard function of Ti in terms of the left truncated Cox proportional hazard model h i (t x i, θ is, t i > a i ) = h 0 (t λ, η 0 ) exp{η 1 Famhist i +η 2 Brstproc i +α m it }, t > a i h 0 (t λ, η 0 ) = λt λ 1 e η 0 : baseline risk function of a Weibull distribution; x i : baseline covariates, First-degree relative with BC (Famhist) and prior breast procedures (Brstproc) with regression coefficients η 1 and η 2 ; α: effect of the individual trajectory of breast density of a woman over their BC risk in terms of the latent BD mean; a i : age over 50 at which woman i enters the screening program thus providing a left truncated time (Uzunogullari et al., 1992); θ is = (λ, η 0, η 1, η 2, α, θ il ) T, with θ il = (β 0, β 1, b 0i, b 1i ) T. 13 / 26

14 The joint model, III 1 Longitudinal submodel: A cumulative logit model for the longitudinal ordinal measures of breast density based on the idea of a continuous latent variable (Lunn et al., 2001; Luo et al., 2014), and y ij = k y ij (γ k 1, γ k ], logit P(y ij > k) = m ij γ k, m it = β 0 + b i0 + (β 1 + b i1)t, 2 Survival submodel: A left-truncated Weibull proportional hazard model for the time-to-bc detection, that incorporates information from the longitudinal process. h i(t x i, θ is, ti > a i) = h 0(t λ, η 0) exp{η 1Famhist i + η 2Brstproc i + α m i(t)} = h 0(t λ, η 0) exp{η 1Famhist i + η 2Brstproc i + α(β 0 + b i0 + (β 1 + b i1)t)}, t 3 Both processes are connected through a shared vector of random effects, which, in the presence of covariates and parameters, endows both processes with conditional independence (Rizopoulos, 2012). 14 / 26

15 Bayes Inference Bayesian Inference: Complete specification of the joint model needs the elicitation of a prior distribution for the subsequent parameters and hyperparameters. Prior independence as a default specification and wide proper prior distributions with the aim of giving all inferential prominence to the data. Normal distributions for the regression coefficients of the longitudinal and survival submodels and the association coefficient. Truncated Normal distributions for the cutpoints of the latent scale, = γ 0 < γ 1 < γ 2 = 0 < γ 3 < γ 4 = selected to provide the same prior probability to each response category. Gamma distribution, Ga(1, 1), for the parameter of the baseline hazard function because it mimics a constant baseline hazard function (Guo et al., ) Uniform hyperdistribution distribution Un(0, 10) for the standard deviations σ 0 and σ 1 of the normal random effects b i0 N(0, σ 0) and b i1 N(0, σ 1). 15 / 26

16 Sección 4 Results 16 / 26

17 Posterior distribution Longitudinal submodel: A cumulative logit model for the longitudinal ordinal measures of breast density based on the idea of a continuous latent variable (Lunn et al., 2001; Luo et al., 2014). y ij = k y ij (γ k 1, γ k ], logit P(y ij > D k ) = m ij γ k, m it = β 0 + b i0 + (β 1 + b i1 )t, Survival submodel: A left-truncated Weibull proportional hazard model for the time-to-bc detection, that incorporates information from the longitudinal process. h i (t x i, θ is, t i > a i ) = h 0 (t λ, η 0 ) exp{η 1 Famhist i + η 2 Brstproc i + α m it } = h 0 (t λ, η 0 ) exp{η 1 Famhist i + η 2 Brstproc i + α(β 0 + b i0 + (β 1 + b i1 )t)}, t > a i Posterior distribution: Approximated by Markov Chain Monte Carlo methods. Mean Sd 2.5 % Median 97.5 % P( > 0 D) β β σ σ γ γ η η η α λ / 26

18 Posterior distribution for the probabilities associated to breast density categories P(y it = a, b, c, d θ, b i, φ) depends on θ and b i. Posterior distribution for the probability associated to each breast density categories for a generic woman i of the same population at time t, π(p(y it = a, b, c, d θ, b i, φ) Data) approximated from f (b i φ) and π(θ, φ Data). Posterior mean 95 % credible interval of π(p(y it = a, b, c, d θ, b i, φ) Data) 0.6 Probability a b c d Age (years) 18 / 26

19 Posterior predictive distribution for the latent breast density Conditional latent breast density (y it θ, b i, φ) Logistic(m it, s = 1). Posterior predictive distribution of the latent breast density of a generic woman i at t years over 50, (y it Data), (50, 55, 60, 65 and 70 in the graphic) Years y* Age (years) Probability a b c d 19 / 26

20 Posterior distribution for the hazard ratios, I Posterior distribution for the hazard ratio (HR): HR of BC diagnosis in women with prior breast procedures versus women without prior breast procedures. HR of BC diagnosis in women with family history of BC versus women without family history of BC. HR of BC diagnosis in women with family history of BC and prior breast procedures versus women without family history of BC and without prior breast procedures var Brstproc Famhist Famhist and Brstproc Hazard Ratio 20 / 26

21 Posterior distribution for the hazard ratios, II Posterior distribution of the HR of a BC diagnosis for a generic woman i with breast density in category k at time t versus this generic woman with breast density in category k at t is NOT POSSIBLE because h i (t x i, θ is, t i h i (t x i, θ is, t i > a i ) = h 0 (t λ, η 0 ) exp{η 1 Famhist i + η 2 Brstproc i + α m it }, YES > a i ) = h 0 (t λ, η 0 ) exp{η 1 Famhist i + η 2 Brstproc i + α y it }, NO We have constructed an ad hoc procedure to represent each value of the breast density at time t in terms of a particular value of the true latent breast density at t ordinal y it latent yit true latent m it (yit = m it + ɛ it ) t = 0 BI-RADS b c d a (1.450, 1.904) (1.959, 3.207) (2.437, 4.684) b (1.351, 1.685) (1.681, 2.460) c (1.244, 1.460) t = 10 BI-RADS b c d a (1.464, 1.937) (2.054, 3.483) (2.569, 5.132) b (1.403, 1.798) (1.754, 2.649) c (1.251, 1.473) 21 / 26

22 Dynamic estimation and prediction, I Bayesian reasoning Estimation of the conditional survival probability that a woman i with longitudinal information Y i at t i,ni and free of a diagnosis of breast cancer at t i,ni is also free at t, P(T i > t θ, b, φ, Y i, x i, T i > t i,ni ), by means of its posterior distribution π(p(t i > t θ, b, φ, Y i, x i, T i > t i,ni ) Data), t > t i,ni that we approach from the posterior distribution π(θ, b, φ Y i, T i > t i,ni, Data). Prediction of the longitudinal breast cancer measurement of a given woman in a scheduled time t ni +1, as the predictive distribution P(y i,tni +1 = a, b, c, d Y i, x i, Data), k = 1, 2, 3, 4 computed from The conditional probability P(y i,tni +1 = a, b, c, d θ, b, φ, Y i, x i, T i > t i,ni ), and The posterior distribution π(θ, b, φ Y i, T i > t i,ni, Data) 22 / 26

23 Dynamic prediction and prediction, II Both posterior distributions π(p(t i > t θ, b, φ, Y i, x i, T i > t i,ni ) Data), t > t i,ni P(y i,tni +1 = a, b, c, d Y i, x i, Data), k = 1, 2, 3, 4 based on π(θ, b, φ Y i, T i > t i,ni, Data) and could be applied to women in the study as well as women not in the study. But we only have an approximate sample from π(θ, b, φ Data). Consequently, if woman i is In the study and Data includes Y i, x i, T i > t i,ni then π(θ, b, φ Data) = π(θ, b, φ Y i, T i > t i,ni, Data) In the study but with new longitudinal and survival information not included in Data then Y i, x i, T i > t i,ni then π(θ, b, φ Data) π(θ, b, φ Y i, T i > t i,ni, Data) Update the posterior distribution with the new information. Sequential MCMC methods in progress. Not in the study with sequential interest we need to update π(θ, b, φ, b i Y i, T i > t i,ni, Data) Update the posterior distribution with the new information. Sequential MCMC methods in progress. 23 / 26

24 Dynamic estimation and prediction, III Posterior distribution: π(p(t i > t θ, b, φ, Y i, x i, T i > t i,ni ) D), t > t i,ni f (y i,tni +1 Y i, x i, D) Posterior mean and 95 % credible interval of the survival probability distribution and posterior mean of the predictive mammogram breast density of woman id17540 (No family history of BC and no prior breast procedures) with a given follow-up which includes all her historical longitudinal mammogram and covariates. 24 / 26

25 Sección 5 Conclusions 25 / 26

26 Conclusions Joint models of longitudinal and survival data are a suitable tool for analyzing the relationship between mammographic breast density and breast cancer risk. Our joint model is a good starting point for learning about the problem but we need to introduce more flexibility and work more on the problem: Non-linear trajectories in the longitudinal mammogram BD modeling. Leave the Weibull baseline risk function and explore other proposals which allow for multimodal or heavy tailed patterns. Assess the performance of the biomarker: Sensitivity, specificity and time-dependent receiver operating characteristic (ROC) curves. Sensitivity of the prior distributions, in particular the hyperpriors for the random effects. Improve our knowledge about latent variables in ordinal models. 26 / 26