Dynamic Malignancy Prediction

• Dynamic malignancy prediction is a technique that integrates time-varying clinical, imaging, and molecular biomarker data to estimate cancer risk progression.
• It employs stochastic and mechanistic models, as well as nonparametric signature extraction, to address sparse and heterogeneous longitudinal data challenges.
• Deep learning and Bayesian joint models further enhance accuracy by incorporating multimodal inputs and individualized risk updates in clinical settings.

Dynamic malignancy prediction refers to the quantitative estimation of malignancy risk or progression through models that utilize time-dependent and longitudinal data. Such models are developed to address clinical scenarios where the temporal evolution of molecular, imaging, or clinical biomarkers is informative of cancer onset or trajectory, and static, single-time-point predictions are fundamentally insufficient. The field has evolved to incorporate mechanistic stochastic models, longitudinal biomarkers, imaging time-series, and advanced statistical feature maps, with rigorous emphasis on performance under data scarcity, heterogeneity, and competing risks.

1. Stochastic and Mechanistic Models for Malignancy Dynamics

A foundational modeling approach utilizes continuous-time Markov processes to capture the birth–death dynamics of malignant and benign cell populations. For example, in the context of circulating tumor DNA (ctDNA) in blood, a malignant clone is represented by a branching process $A_t$ with birth rate $b$ and death rate $d$, yielding a net growth rate $r = b - d > 0$. Apoptotic cell death produces ctDNA fragments with probability $q_d$, creating an observable process $C^A_t$ for ctDNA (Vaucher et al., 10 Jun 2025).

Benign background shedding is typically modeled via a death-with-immigration process, yielding Poisson-distributed ctDNA concentrations. These models are advantageous for noninvasive cancer monitoring, leveraging a handful of blood samples to characterize the underlying tumor dynamics without the need for explicit parameter estimation in regimes with sparse temporal sampling.

2. Feature Extraction and Nonparametric Representation

Recent work recognizes the limitations of direct parameter inference (e.g., estimating $b,d,\lambda$) when only a few time points are available. Signature theory is introduced to systematically summarize the "shape" of multivariate time series (e.g., $(t, C_t)$) with a hierarchy of pathwise iterated integrals $S_{0,T}^{i_1 \cdots i_k}(X)$. These signatures are reparametrization-invariant and encode nonlinear time–biomarker interactions directly, yielding efficient feature maps suitable for irregularly sampled, low-data regimes (Vaucher et al., 10 Jun 2025).

For instance, key order-1 and order-2 signature statistics (such as $S^2 = C_T - C_0$ and the Levy area $LA$) admit exact null distributions (Skellam laws) under the benign hypothesis, allowing construction of principled hypothesis tests and computation of p-values with controlled false discovery rates.

3. Joint Modeling and Landmarking Strategies

Dynamic malignancy risk is also estimated using joint models that link longitudinal biomarker trajectories with time-to-event outcomes. A typical joint model specifies a nonlinear mixed-effects submodel for the biomarker (e.g., $y_i(t) = m_i(t) + \varepsilon_i(t)$, with random effects), and a proportional hazards survival submodel with time-dependent hazard functions $h_i(t)$ modulated by current, slope, or cumulative history of the biomarker. Full likelihood integration over random effects supports patient-specific survival probability estimation at arbitrary time points (Rizopoulos et al., 2013).

Landmarking is an alternative, simpler method: at each "landmark" time $t$, survival predictions are updated using Cox models incorporating the most recent biomarker measurement or summary. These methods have been extended to competing risks, e.g., proportional subdistribution hazards landmark supermodels, which allow for dynamic prediction of cumulative incidence functions (CIFs) while incorporating time-varying covariates and non-proportional effects (Liu et al., 2019).

4. Deep Learning Architectures for Longitudinal Imaging and Multimodal Biomarkers

Deep architectures extend dynamic malignancy prediction to high-dimensional imaging and multimodal biomarker inputs. In pulmonary nodule follow-up, pipelines integrate 3D convolutional networks for detection, hierarchical probabilistic U-Nets for uncertainty-aware growth quantification, and two-stream temporal classifiers to combine appearance and change signals, yielding improved malignancy classification performance (AUC up to 0.910) and more accurate uncertainty calibration compared to single-timepoint or diameter-only methods (Rafael-Palou et al., 2021).

Time-agnostic, multitask diffusion networks have been demonstrated for brain tumor progression, enabling pixel-wise probabilistic tumor evolution mapping across arbitrary future temporal horizons. These models leverage signed distance fields (SDFs) to capture spatial uncertainty, pretrained deformation modules for temporal consistency, RT-dose weighted focal losses, and targeted synthetic data augmentation to overcome sparse follow-up and missing modality challenges (Kebaili et al., 13 Sep 2025).

Longitudinal mammography-based risk prediction frameworks have evolved toward recurrent deep state-space models (e.g., Vision Mamba RNNs) with LSTM-like gating and asymmetry-aware modules that explicitly track spatial and temporal tissue differences, achieving superior long-term cancer risk forecasting and robustness in high-density cases (Sun et al., 20 Jun 2025).

5. Bayesian Hierarchical and Mechanistic Models

Dynamic prediction under treatment, especially in hematologic or solid malignancies, increasingly utilizes hierarchical Bayesian models that marry mechanistic disease progression equations with flexible, covariate-dependent random effects. For example, Bayesian joint models for multiple nonlinear biomarkers and cause-specific hazards enable individualized prediction of death or treatment transitions in multiple myeloma, integrating nonlinear bi-exponential trajectories and competing risks, estimated via full MCMC or "corrected" two-stage strategies for computational efficiency (Alvares et al., 2024).

In simulated tumor biomarker dynamics, subpopulation models parameterized by growth and decay rates, resistant fractions, and initial burden ($\rho^r, \rho^s, \pi^r, M^0$), when combined with Bayesian neural network mappings from baseline covariates, outperform traditional linear models—especially when covariate interactions influence resistance and regrowth dynamics. Principled uncertainty quantification and mechanistic interpretability are preserved (Myklebust et al., 2024).

6. Validation, Calibration, and Clinical Performance

Performance evaluation in dynamic malignancy prediction encompasses discrimination metrics (time-dependent AUC, dynamic concordance index), calibration statistics (Brier score, observed-to-expected ratios), and interval coverage—often estimated via bootstrapping or Monte Carlo integration over model posteriors. Multiple frameworks provide consistent, robust estimation under sparse, noisy, and competing-risk conditions, sustaining high accuracy, precision, and recall in both simulated and real world datasets for ctDNA trajectories (Vaucher et al., 10 Jun 2025), imaging follow-up (Rafael-Palou et al., 2021, Kebaili et al., 13 Sep 2025, Sun et al., 20 Jun 2025), and joint longitudinal outcomes (Alvares et al., 2024).

7. Limitations, Practical Recommendations, and Emerging Directions

Dynamic malignancy prediction frameworks confront inherent challenges: limited sampling, uncertainty in biomarker kinetics (e.g., ctDNA elimination), potential model misspecification (e.g., ignoring necrosis or active secretion), and computational demands (in Bayesian neural networks or high-dimensional deep learning). Model validation is often restricted to simulated or modestly sized real cohorts.

Emerging best practices include:

• Utilizing nonparametric or invariant feature extraction (signatures, SDFs) for sparse data (Vaucher et al., 10 Jun 2025, Kebaili et al., 13 Sep 2025).
• Integrating uncertainty estimates (HPU-based diameter distributions, Bayesian credible intervals) for risk stratification (Rafael-Palou et al., 2021, Myklebust et al., 2024).
• Modularizing architectures to exploit prior knowledge (frozen encoders, pretrained deformation priors) while training minimal temporal/asymmetry heads (Sun et al., 20 Jun 2025, Kebaili et al., 13 Sep 2025).
• Employing multi-model validation spanning calibration, discrimination, and interval coverage statistics (Rizopoulos et al., 2013, Liu et al., 2019, Alvares et al., 2024).
• Real-time updating of individualized predictions as new longitudinal data accrue (Alvares et al., 2024).

Future developments will involve extension to multivariate and multimodal inputs, increased external validation, adoption of higher-order nonparametric features, and advances in computationally efficient Bayesian inference for integration with clinical workflows.