QuanTIM Webinar, SESSTIM research unit, Marseille, France
June 16, 2023

Background & Motivation

PCa Active Surveillance

  • To avoid over-treatment, men with low grade prostate cancer are advised active surveillance


  • Cancer progression is tracked via:
    • Prostate-specific antigen measurements
    • Digital rectal examination
    • MRI
    • Biopsies

PCa Active Surveillance (cont’d)

  • Treatment is advised when cancer progression is observed
    • typically via biopsies when Gleason Score \(\geq 7\)




Frequency of Biopsies

PCa Active Surveillance (cont’d)

  • Two dimensions
    • Number of biopsies
    • Delay in finding progression


  • Delay: We want to find progression asap


  • Number of biopsies: high burden
    • painful, cause complications, expensive

Biopsies Schedules


  • Annual Biopsies
    • focus on minimizing delay
    • many unnecessary biopsies for patients who progress slow

Biopsies Schedules (cont’d)


  • Less Frequent Biopsies - PRIAS
    • every 3 years or
    • annually if PSA doubling time < 10 (try to find faster progressions)
    • still unnecessary biopsies for patients who progress slow

Biopsies Schedules (cont’d)

  • unnecessary biopsies \(\Rightarrow\) Low compliance
    • effectiveness of AS is compromised




Considerable room to improve biopsy scheduling

A New Approach: Personalized Scheduling

A New Approach

  • Scheduling based on individualized risk predictions
    • Progression rate is not only different between patients but also dynamically changes over time for the same patient


  • Risk predictions based upon
    • All available PSA (ng/mL) measurements
    • All available DRE (T1c / above T1c) measurements
    • Time and results of previous biopsies

A New Approach (cont’d)

A New Approach (cont’d)

A New Approach (cont’d)



How to better plan biopsies?


  • In steps:
    • How the longitudinal PSA & DRE are related to Gleason reclassification?
    • How to combine previous PSA & DRE measurements and biopsies to predict reclassification?
    • When to plan the next biopsy?

Modeling Framework

Time-varying Covariates

  • To answer these questions we need to link
    • the time to Gleason reclassification (survival outcome)
    • the PSA measurements (longitudinal continuous outcome)
    • the DRE measurements (longitudinal binary outcome)


  • Biomarkers are endogenous time-varying covariates
    • their future path depends on previous events
    • standard time-varying Cox model not appropriate

Time-varying Covariates (cont’d)



To account for endogeneity we use the framework of



Joint Models for Longitudinal & Survival Data

The Basic Joint Model

DynPreds

The Basic Joint Model (cont’d)

  • We need some notation
    • \(T_i^*\) the true reclassification time
    • \(T_i^L\) last biopsy time point Gleason Score was \(< 7\)
    • \(T_i^R\) first biopsy time point Gleason Score was \(\geq 7\)
    • \(T_i^R = \infty\) for patients who haven’t been reclassified yet
    • \(\mathbf y_{i1}\) vector of longitudinal PSA measurements
    • \(\mathcal Y_{i1}(t) = \{y_{i1}(s), 0 \leq s < t\}\)
    • \(\mathbf y_{i2}\) vector of longitudinal DRE measurements
    • \(\mathcal Y_{i2}(t) = \{y_{i2}(s), 0 \leq s < t\}\)

The Basic Joint Model (cont’d)

\[\left \{ \begin{array}{ccl} h_i(t) & = & h_0(t) \exp \{\mathbf \gamma^\top \mathbf w_i + \alpha_1 {\color{red} \eta_{i1}(t)} + \alpha_2 {\color{blue} \eta_{i2}(t)}\}\\&&\\ y_{i1}(t) & = & {\color{red} \eta_{i1}(t)} + \varepsilon_i(t)\\ & = & \mathbf x_{i1}^\top(t) \mathbf \beta_1 + \mathbf z_{i1}^\top(t) \mathbf b_{i1} + \varepsilon_i(t)\\&&\\ \log\frac{\Pr\{y_{i2}(t) = 1\}}{1 - \Pr\{y_{i2}(t) = 1\}} & = & {\color{blue} \eta_{i2}(t)}\\ & = & \mathbf x_{i2}^\top(t) \mathbf \beta_2 + \mathbf z_{i2}^\top(t) \mathbf b_{i2}\\&&\\ \mathbf \{b_{i1}, b_{i2}\} \sim \mathcal N(\mathbf 0, \mathbf D), & & \varepsilon_i(t) \sim \mathcal N(0, \sigma^2) \end{array} \right.\]

The Basic Joint Model (cont’d)


  • The longitudinal and survival outcomes are jointly modeled \[\begin{eqnarray} p(y_{i1}, y_{i2}, T_i^L, T_i^R) & = & \int p(y_{i1} \mid {\color{red} b_{i1}}) \; p(y_{i2} \mid {\color{red} b_{i2}}) \times \\ && \quad \quad \left\{S(T_i^L \mid {\color{red} b_i}) - S(T_i^R \mid {\color{red} b_i})\right\} p({\color{red} b_i}) \; d{\color{red} b_i}\\ \end{eqnarray}\]
    • the random effects \({\color{red} b_i}\) explain the interdependencies

Functional Form


  • PSA velocity

    • fast increasing PSA indicative of progression


    \[h_i(t) = h_0(t) \exp \{\mathbf \gamma^\top \mathbf w_i + \alpha_1 {\color{red} \eta_{i1}(t)} + \alpha_2 {\color{blue} \eta_{i1}'(t)} + \alpha_3 {\color{red} \eta_{i2}(t)}\}\]
    where \({\color{blue} \eta_{i1}'(t)} = \frac{d}{dt} \eta_{i1}(t)\)

Functional Form (cont’d)

TD_slopes

Personalizing the Biopsy Schedules

Risk of Progression

  • Using the fitted joint model we can calculate the cumulative risk of progression \[\pi_j(u \mid t, v) = \Pr \bigl \{ T_j^* \leq u \mid T_j^* \geq t, \mathcal Y_{j1}(v), \mathcal Y_{j2}(v) \bigr \}\]


  • \(t\) time of last biopsy
  • \(v\) time of current visit, \(v \geq t\)
  • \(u\) future time, \(u \geq t\)
  • \(\mathcal Y_{j1}(v)\) & \(\mathcal Y_{j2}(v)\) available PSA & DRE measurements up to current visit

Risk of Progression (cont’d)

Personalized Schedule

  • Patients come back every 6 months for PSA & DRE measurements
    • at these occasions we want to decide for a biopsy


  • In general, we consider decisions at a fixed schedule \[\begin{array}{l} s_1, \ldots, s_N\\ s_1 = v\\ s_N = h \end{array}\]

Personalized Schedule (cont’d)

Personalized Schedule (cont’d)

  • Simple decision rule: We do a biopsy at \(s_n\) if

    \[\pi(s_n \mid t_n, v) \geq \kappa_n\]
  • where
    • \(\kappa_n\) a threshold at \(s_n\)
    • \(t_n\) time of last biopsy before \(s_n\)

Personalized Schedule (cont’d)

Personalized Schedule (cont’d)

  • The key question is



How do we select \({\color{red} \kappa_n}\)?

Personalized Schedule (cont’d)

  • We consider two relevant quantities
    • the number of biopsies
    • the delay in finding progression



Ideally, we would like to just do one biopsy at exactly the time point of progression

Personalized Schedule (cont’d)

  • For different thresholds \(\kappa_n\) we would obtain different number of biopsies and different delays…


  • For a specific threshold \(\kappa^*\) we can calculate
    • how many times a biopsy will be performed in the future

Personalized Schedule (cont’d)

Personalized Schedule (cont’d)

  • The times when biopsies are performed \[t_n = \left\{ \begin{array}{l} t_{n-1} : \pi_j(s_n \mid t_{n - 1}, v) < \kappa^*\\\\ s_n : \pi_j(s_n \mid t_{n - 1}, v) \geq \kappa^*\\ \end{array} \right.\]


  • The expected number of biopsies will be \[\mathcal N_j(\kappa^*) = \sum_{n = 1}^N \mbox{I}\{\pi_j(s_n \mid t_n, v) \geq \kappa^*\} \times \{1 - \pi_j(t_{n-1} \mid t, v)\}\]

Personalized Schedule (cont’d)

  • For a specific threshold \(\kappa^*\) we can calculate
    • the expected delay \[\begin{array}{lcl} \mathcal D_j(\kappa^*) & = & \sum\limits_{n = 1}^N \bigl \{ t_n - E(T_j^* \mid t_{n-1} \leq T_j^* \leq t_n) \bigr \} \times\\&&\\ && \quad \quad \Pr \bigl \{ t_{n-1} \leq T_j^* \leq t_n \mid T_j^* > t, \mathcal Y_{j1}(v), \mathcal Y_{j2}(v) \bigr \} \end{array}\]

Personalized Schedule (cont’d)



  • For different \(\kappa\)’s we construct the two-dimensional space of expected number of biopsies and expected delays

Personalized Schedule (cont’d)

Personalized Schedule (cont’d)


  • If we consider that the delay & the number of biopsies are equally important
    • we can select the \(\kappa_n\) that is closest to the optimal schedule \[{\color{red} \kappa_n^{opt}} = \mbox{argmin}_{\kappa} \sqrt{ \bigl \{ \mathcal N_j(\kappa) - 1 \bigr \}^2 + \mathcal D_j(\kappa)^2}\]

Personalized Schedule (cont’d)


  • Othrewise,
    • we may also select a clinically acceptable delay, and
    • select \({\color{red} \kappa_n^{opt}}\) the \(\kappa\) that minimizes the expected number of biopsies

Discussion

Some Considerations

  • Calibration
    • calculation of expected delay and number of biopsies require a well-calibrated model


  • Schedules become more personalized the better biomarkers distinguish patients
    • consider more biomarkers, e.g., for postate cancer MRI

Resources