Dynamic Prediction of Clinical Endpoints in Survival Analysis With Machine and Deep Learning: A Systematic Review and Regulatory Perspective

ABSTRACT

Accurate prediction of time‐to‐event outcomes is essential for clinical decision‐making and model‐informed drug development (MIDD). Conventional survival models may inadequately represent complex longitudinal patient trajectories, time‐dependent treatments, and evolving biomarker or pharmacokinetic/pharmacodynamic (PK/PD) profiles. Recent advances in machine learning (ML) and deep learning (DL) have enabled dynamic survival models that update individualized risk predictions as new longitudinal information becomes available. This report systematically reviews dynamic survival modeling frameworks, including extensions of Cox proportional hazards (CPH) with time‐varying covariates, discrete‐time landmarking strategies, random survival forests (RSF), and deep learning approaches such as DeepHit, Dynamic‐DeepHit, and neural ordinary differential equations (ODEs). Applications span sepsis detection, cancer prognosis, and neurodegenerative disease progression. Dynamic models generally achieved discrimination comparable to or better than static Cox models when longitudinal data were available; however, calibration was rarely reported, external validation was uncommon, and integration of PK/PD features into dynamic survival frameworks remains underexplored. Integrating exposure–response modeling with dynamic ML/DL may advance precision dosing, support adaptive trial design, and enhance translational relevance in pharmaceutical research. Emerging regulatory frameworks, including ICH M15 and FDA draft guidance on artificial intelligence in drug development, further emphasize clearly defined contexts of use, risk‐based model assessment, credibility, and transparent reporting. Dynamic prediction at the intersection of ML/DL, PK/PD, and regulatory science represents a promising path toward clinically deployable models to inform drug development and regulatory decision‐making.

1. Introduction

Accurate prediction of time‐to‐event outcomes, such as progression‐free survival, overall survival, and time to dose‐limiting toxicity, is central to clinical development, informing dose selection, trial enrichment, and benefit–risk assessment. In oncology drug development, traditional statistical models remain widely used. Logistic regression is commonly applied for the characterization of binary safety endpoints such as dose‐limiting toxicities, while time‐to‐event models continue to support progression‐free (PFS) and overall survival (OS) analyses. These established frameworks provide context for interpreting emerging dynamic survival models. However, conventional survival models, such as Cox proportional hazards or parametric time‐to‐event models, provide interpretable hazard ratios but assume proportional effects and often ignore the rich longitudinal information embedded in drug exposure (pharmacokinetic/pharmacodynamic, PK/PD) and biomarker trajectories. As a result, they can be ill‐suited for the high‐dimensional, irregularly sampled data characteristic of modern trials [1].

Recent machine learning (ML) and deep learning (DL) methods overcome many of these limitations by flexibly modeling nonlinear relationships and leveraging the full history of time‐varying covariates. Recurrent neural networks, long short‐term memory (LSTM) networks, temporal convolutional networks, attention‐based transformers, and neural ordinary differential equation (Neural‐ODE) models can ingest continuous streams of dosing, concentration, and biomarker data and produce dynamic hazard functions that update as new information arrives. In critical care and oncology, such models achieve higher discrimination and have attracted increasing attention for real‐time monitoring and early warning systems [2, 3]. This positions dynamic ML/DL survival models as valuable tools for model‐informed precision dosing and adaptive trial design.

In traditional pharmacometrics practice, exposure–response models are typically refitted in full whenever new data become available. This process, while rigorous, can be time‐consuming, computationally intensive, and poorly suited for real‐time clinical decision support. This limitation underscores the clinical and operational value of dynamic prediction frameworks. Unlike traditional exposure–response modeling workflows that require full model refitting, dynamic ML/DL models can update predictions immediately as new dosing, concentration, or biomarker information becomes available. This incremental updating is more efficient, better reflects evolving patient status, and supports real‐time therapeutic decision‐making—capabilities increasingly important for dose titration, early safety monitoring, and adaptive trial conduct. Dynamic models incorporate repeated measurements, laboratory values, vital signs, imaging, and importantly, PK/PD data, so that risk predictions reflect the current physiologic and treatment status; static models cannot adapt to changes over time [4, 5, 6]. Incorporating exposure histories (e.g., drug concentrations, cycle area under the curve, trough levels) and downstream pharmaco‐dynamic markers allows individualized risk to be updated as dosing changes, concentrations fluctuate, and biomarkers respond. Such PK/PD integration has been shown to improve discrimination and align predictions with titration decision points [7, 8, 9]. This integration aligns predictions with real decision points in titration and therapeutic drug monitoring and makes them especially valuable for pharmaceutical stakeholders.

Several recent reviews have mapped portions of this landscape (Table 1). Tran et al. [10] emphasized multi‐omics cancer survival prediction, contrasting interpretable regression‐based methods (e.g., IPF‐LASSO, Priority‐Lasso) with advanced neural approaches such as SALMON, GDP, and MiNet, and highlighting discrete‐time models for time‐varying covariates (e.g., TF‐Loghazard Net). Wiegrebe et al. [11] offered a comprehensive survey of deep learning methods, spanning FFNNs, RNNs, CNNs, autoencoders, and neural ordinary differential equations (Neural‐ODEs). Smith et al. [12] conducted simulation‐based comparisons of 29 methods, including Cox models, RSF, boosting, SVMs, and hybrid learners, to systematically assess performance under censoring and nonproportional hazards. Kantidakis et al. [2] appraised neural networks for survival prediction, focusing on designs that encode time as input or as discrete intervals, with applications in oncology and cardiology. Kaur et al. [1] surveyed ML methods for cancer prognosis across SEER and other datasets, covering decision trees, ensembles, ANNs, and hybrid models for prognosis and treatment guidance. Cascarano et al. [13], reviewed ML/DL strategies for longitudinal biomedical data, including temporal feature engineering, trajectory clustering, and RNN‐based time‐series modeling, while noting challenges of data irregularity and missingness. Most recently, Huang et al. [14] conducted a scoping review of ML applications to survival outcomes using real‐world data. However, none of these prior reviews focused specifically on dynamic, time‐updating prediction frameworks or on their integration with pharmacokinetic/pharmacodynamic (PK/PD) exposure data—the central themes of the present review. This gap, together with the pharmacometric and regulatory framing adopted here, motivates the present synthesis.

TABLE 1.

Summary of review papers.

First author (year)	Topic	ML/DL methods and usage
Tran, D (2025) [10]	Cancer survival prediction using multi‐omics and clinical data	•Summarizes regression‐based techniques to model survival outcomes with interpretable coefficients (e.g., IPF‐LASSO, M2EFM, Priority‐Lasso) •Apply neural networks to capture complex patterns in multi‐omics and clinical data (e.g., SALMON, GDP, MiNet, SurvivalNet, SAE, CSAE) •Highlights time‐varying covariates and model survival in discrete time (e.g., TF‐Loghazard Net, TF‐ESN, MDNNMD)
Wiegrebe, S (2024) [11]	Systematic review of deep learning methods for survival analysis	•Reviews DL‐based survival methods across various architectures: FFNNs, RNNs, CNNs, autoencoders, and neural ODEs •Highlights models designed for complex survival tasks including competing risks (e.g., DeepHit, DeepComp, CRESA), multi‐state outcomes (e.g., survNode, DeepPAMM), and recurrent events (e.g., CRESA, DeepPAMM) •Discusses handling of time‐varying features and effects using RNNs and Transformer‐based models (e.g., Dynamic‐DeepHit, RDSM, TransformerJM) •Identifies methods tailored for high‐dimensional and multimodal data (e.g., SALMON, MCAP, PAGE‐Net, MultiSurv), integrating omics, imaging, and clinical features
Smith, H (2022) [12]	Simulation‐based comparison of statistical and ML methods for time‐to‐event risk prediction	•Compares 29 methods across 10 simulation studies: 4 statistical (mostly Cox‐based), 22 ML (e.g., RSF, neural networks, boosting, SVM), and 3 hybrid approaches •ML methods include DeepSurv, IPCW‐wSVM, MKNN‐KM, CUDLs, ELMCoxBAR, and Super Learner variant •Hybrid methods combine ML with statistical modeling (e.g., boosting with Cox, Super Learner, MKNN‐KM) •Application scenario: evaluating prognostic model performance using simulated survival data with varying assumptions (e.g., censoring, non‐proportional hazards)
Kantidakis, G. (2022) [2]	Review and appraisal of neural networks for survival prediction using prognostic factors	Uses artificial neural networks (ANNs) for survival prediction, employing two modeling strategies: time as input with a single output node, or multiple output nodes for discrete time intervals Applied to clinical prognostic data across oncology and cardiovascular domains to estimate survival probabilities and risk scores
Kaur, I. (2022) [1]	Survey of ML techniques for cancer survival prediction and prognosis	•Uses decision trees, ANNs, ensemble methods, and hybrid models for survival prediction, risk stratification, and treatment recommendation •ML models applied to SEER and clinical datasets across cancer types (e.g., breast, oral, lung, rectal) to capture non‐linear patterns and improve prognostic accuracy
Cascarano, A. et al. (2023) [13]	General review of longitudinal biomedical data analysis using ML/DL	•Discusses using temporal features within models like SVM and tree ensembles, and recurrent neural networks (e.g., RNNs) applied directly to time‐series data •Covers clustering approaches for trajectory stratification, though the primary focus is on static outcome prediction from longitudinal features •Evaluates strengths and limitations of each strategy, e.g., interpretability of summary features vs. expressivity of RNNs, and challenges posed by data irregularity and missingness
Huang, Y. et al. (2023) [14]	Application of machine learning in predicting survival outcomes involving real‐world data	•Random survival forests and neural networks, are mainly used for real‐world data to predict survival outcomes such as disease prognosis or clinical events in the oncology

Despite the promise of ML/DL‐enabled dynamic survival modeling, several challenges remain. The literature is fragmented across many modeling frameworks, external validation is limited, and the integration of PK/PD information is still nascent [7]. Many ML/DL architectures are also data‐hungry: training and properly validating these models typically requires large, high‐quality longitudinal datasets, which can be difficult to obtain in rare diseases, small trials, or settings with limited follow‐up, and this constraint should be weighed against potential gains in predictive performance [15]. In addition, regulatory expectations for the use of advanced modeling in drug development are evolving. Draft guidance such as ICH M15 and the FDA's considerations for AI highlight that credibility depends on clear articulation of context of use, risk–impact assessment, and transparent reporting.

This review synthesizes recent advances in dynamic survival modeling for clinical and pharmacometric applications, classifies ML/DL approaches by temporal design and update frequency, and evaluates evidence for incorporating drug exposure. By highlighting methodological gaps, regulatory perspectives, and opportunities, we aim to guide the development of robust, interpretable dynamic models that can better support pharmaceutical decision‐making.

2. Methods

2.1. Search Strategy

We conducted a comprehensive search to identify studies applying artificial intelligence and machine learning to prognostic modeling in human clinical settings with direct relevance to pharmaceutical development. PubMed was the primary source, queried with the prespecified strategy—((“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network”) AND (“prediction” OR prognostic OR “risk model”) AND (“clinical” OR “disease”) AND (“survival”)) — and was complemented by targeted supplementary sources (trial registers and adjacent outlets) to ensure capture of drug‐development–relevant work that may be variably indexed in PubMed. This query retrieved a broad set of ML/DL survival‐prediction studies spanning both static and dynamic approaches (see Table 2), with studies incorporating longitudinal biomarker or covariate data prioritized during screening. We retained this broader set to contextualize the dynamic, time‐updating models that are the focus of this review (Sections 3.2, 3.4). In the identification phase, the flow diagram records 6101 items from PubMed and 29 from registers. Before screening, records were removed for predefined reasons lack of eligible full text (n = 64), language other than English (n = 22), or publication date > 10 years prior (n = 159), yielding 5855 records for title/abstract screening. During screening, we excluded obviously out‐of‐scope items by article type or population preprints (n = 41) and animal studies (n = 2086), leaving 3728 reports for eligibility assessment. At full‐text eligibility, we removed reports not sufficiently related to the review topic (n = 3186), which yielded a set of 542 records meeting article‐type and AI/ML‐relevance thresholds (clinical study, methodological paper, or review/systematic review using AI/ML). Applying our final eligibility rules—requiring an AI/ML model developed or applied to human‐study data with survival/time‐to‐event endpoints, quantitative performance reporting, and clear pharmaceutical/drug‐development relevance (e.g., PFS/OS, toxicity, exposure–response, patient stratification or dose optimization)—we excluded studies not focused on the searched aims (n = 467) and retained 75 studies for inclusion. The emphasis throughout eligibility was on models that inform drug‐development decisions: analyses embedded in or directly supporting trials, survival endpoints germane to benefit–risk assessment, integration of PK/PD or exposure measures, and use‐cases that guide therapy selection or trial design. Approximately thirty of the included papers originated outside PubMed (reflected in the register/other‐source stream of the diagram). Their inclusion was intentional and necessary for completeness: clinically grounded ML survival studies tied to trials, sponsor collaborations, engineering venues, or PK/PD analyses are sometimes indexed later or inconsistently in PubMed; backward/forward citation chasing from pharma‐relevant seed papers also surfaced eligible reports that the strict PubMed string did not retrieve at the time of search. Incorporating these sources preserves coverage of decision‐critical AI prognostic evidence across the drug‐development ecosystem. This PubMed‐primary search, supplemented by targeted citation chasing and domain expertise, was designed to prioritize pharmaceutical‐development relevance over exhaustive multi‐database coverage; search scope and reproducibility are addressed as limitations in Section 4.1. Figure 1 shows the PRISMA diagram, which reflects the overall search and selection process.

TABLE 2.

AI algorithms in non‐review papers.

Algorithm family	No. of studies	Temporal design		Recurrent event?	Representative studies
		Dynamic	Disc./Cont.
Cox‐based
CPH	25	No	Cont.	No	[16, 17]
Time‐dependent Cox a	2	Yes	Cont.	No	[18, 19]
Penalized Cox b	6	No	Cont.	No	[20]
Tree‐based
RSF	28	Yes	Cont.	No	[21, 22, 23, 24, 25]
Boosted trees c	15	No	Cont.	No	[20, 26]
Neural networks
DeepSurv	11	No	Cont.	No	[3, 27, 28, 29, 30]
LSTM/RNN	6	Yes	Disc.	No	[31, 32]
Deep‐MTLR	3	No	Disc.	No	[32, 33, 34]
Dynamic‐DeepHit	5	Yes	Disc.	No	[35]
Neural ODE survival	4	Yes	Cont.	No	[36]
Transformer‐based d	2	Yes	Cont.	Yes	[9, 37]
Autoencoder‐based	1	No	Cont.	No	[38]
MLP/ANN	2	No	Disc.	No	[39, 40]
Surv‐GCNN	1	No	Cont.	No	[41]
Ensemble e	3	Yes	Disc.	No	[6]
Other
SVM	10	No	Disc.	No	[42, 43, 44, 45]
Bayesian model	1	Yes	Disc.	Yes	[46]
Joint model	2	Yes	Cont.	No	[47, 48]

Abbreviations: ANN, artificial neural network; Cont., continuous time; Disc., discrete time; MLP, multilayer perceptron; Surv‐GCNN, survival graph convolutional neural network.

^a Cox with time‐varying covariates (non‐PH accommodated).

^b LASSO/elastic‐net/ridge Cox; includes CoxNet.

^c AdaBoost, GBM/GBS/GBRTS, XGBoost, GLMBoost, CoxBoost.

^d Includes TransformerJM/TransformerLSR.

^e Super Learner, AutoPrognosis.

FIGURE 1.

PRISMA flow diagram of study identification, screening, and inclusion (see Section 2.1, Search Strategy, for the corresponding counts at each stage).

3. Results

In total, our review included 75 papers. Of these, seven were review papers synthesizing methodological advances and applications of survival analysis, 56 were methodology papers that proposed, extended, or benchmarked survival models, and 12 were application‐focused studies applying existing models to specific clinical datasets. This distribution reflects the predominance of methodological innovation in the field, supported by a smaller but important body of reviews and application studies that contextualize these developments within clinical and translational research. For example, ML‐based survival models have been applied to predict COVID‐19 mortality from routine laboratory parameters [49], and graph neural networks (GNN) have been used to enhance ovarian cancer survival prediction [50]. Table 2 summarizes the models presented in the non‐review paper, highlighting their key properties, including their suitability for dynamic settings and temporal design considerations. Table 3 provides details of studies that employed dynamic survival prediction, including the number of patients, observation windows, prediction windows, and the AI models utilized.

TABLE 3.

Summary of dynamical survival models.

Objective (Year)	ML/DL models	Source (No. subjects, observation window)	Prediction time	Validation type	Performance a
Real‐time risk prediction in ICU patients (2025) [32]	TBAL (extension of LSTM)	MIMIC‐IV (58,323 stays) + eICU‐CRD (118,021 stays); 176,344 total ICU stays	ICU mortality risk; landmark: every 4 h; window: 24 h (static benchmarks: 12 h–7 days)	7:2:1 patient‐level holdout split; cross‐database external validation (MIMIC‐IV ↔ eICU‐CRD)	AUROC: 0.92–0.96 (internal); 0.76–0.81 (cross‐database)
Explainable mortality prediction in ICU using high frequency EHR (2020) [51]	NN + LSTM	4 Copenhagen ICUs; 11,492 patients/14,190 admissions (2011–2016); external: 5827 patients (2012–2017)	90‐day mortality risk; landmark: updated hourly; window: 90 days	80/20 internal holdout split + external validation	AUROC: 0.73–0.88 (internal); 0.75–0.83 (external)
Dynamic survival analysis with competing risks (Dynamic‐DeepHit, 2020) [35]	Dynamic‐DeepHit	UK Cystic Fibrosis Registry; 5883 adult patients with annual follow‐ups (2009–2015)	Mortality risk (competing: respiratory failure vs. other‐cause death); landmark: each annual visit (ages 30, 40, 50 years); window: 1–5 years	5× repeated 80/20 train/test split (20% of training reserved for hyperparameter validation)	4.36%/9.67% gain over benchmark (respiratory/other‐cause mortality). C_k(t, Δt)
Risk prediction using RF‐SLAM (2019) [52]	RF‐SLAM ensemble	382 patients, 75 SCA (sudden cardiac arrest) events, 8‐year follow‐up (three sites)	SCA risk; landmark: each 6‐month interval; window: 6 months	OOB (out‐of‐bag) error estimation + bootstrap	Time‐varying AUC(t); Spiegelhalter's z‐statistic (calibration)
Landmarking and ensemble‐based survival prediction (2020) [6]	Landmark + Super Learner	UK CF Registry; 6363 CF patients; 43,880 annual review records; 962 composite events (death or lung transplant)	Mortality/lung transplant risk; landmark: annual ages 20–50 years; window: 2 and 5 years	10‐fold cross‐validation (CV) on training data + holdout test from 18 randomly selected CF centers (20% of records, 14% of patients)	C‐index: > 0.80 (2‐years); 0.85–0.90 (5‐years, ages 20–40). Brier Score: 0.03–0.07
Ovarian cancer survival prediction with temporal analysis and GNN (2024) [50]	Temporal GNN	Ovarian Cancer Database (OCD) + SEER (CDC); N not reported	Ovarian cancer survival risk (binary); landmark: none; window: 3 and 6 months	80/10/10 train/validation/test split; k‐fold CV for hyperparameter tuning	Precision: 87%–98%; accuracy: 87%–96%; recall: 84%–96%
Alzheimer's progression prediction using multivariate outcomes (2019) [47]	MFPCACox and Joint model	ADNI‐1; 355 MCI patients (180 ad events, avg. 3.2‐years follow‐up)	AD conversion risk; landmark: 0.5, 1, 1.5, 2 years; window: 0.5 and 1 years	10‐fold CV on real data; simulation: 100 datasets (200 training/100 validation subjects)	Time‐dependent AUC: 0.71–0.88; Brier score < 0.15
Localized prostate cancer survival analysis with deep learning (2022) [53]	RDSM, DSM	VA Health System; 112,276 localized PC patients (2001–2017); ~10‐years max follow‐up	Mortality risk; landmark: diagnosis (t = 0); window: 2, 5, 10 years	80/20 random train/test split + leave‐one‐US‐region‐out CV (proxy for external validation)	C_td: 0.76–0.85 (2–10 years, composite); 0.74–0.84 (leave‐one‐region‐out)
Health status and treatment effect prediction for CLL using dynamic Bayesian networks (2022) [46]	Dynamic Bayesian Network	Knowledge‐based (no patient‐level training data); validated vs. SEER (US, 18 registries) and EUROCARE‐5 (EU, 99 registries); CLL diagnosed 2000–2007, 60‐month follow‐up	CLL survival risk; landmark: none (6‐month state updates); window: 5 and 10 years	Comparing simulated survival curves to SEER and EUROCARE‐5 registry benchmarks	Mean abs. diff.: TEN 1.6% (EUROCARE‐5), 7.8% (SEER); HSN 11.4%/20.9%
Sepsis prediction with DeepAISE (2021) [54]	DeepAISE	Emory ICU cohort: 25,820 patients (2014–2018, 2 hospitals); external: MIMIC‐III ICU database (2001–2012, Beth Israel Deaconess MC)	Sepsis onset risk; landmark: 4 h post‐ICU admission; window: 4, 6, 8, 10, 12 h ahead	70/10/20 split on Emory cohort (train/hyperparameter/test) + external validation on MIMIC‐III	AUROC: 0.88–0.90 (internal, 4–12 h); 0.84–0.87 (external)
Individual dynamic prediction via landmark approach (2022) [5]	RSF, Cox and sPLS‐DR + Super Learner	PBC2 clinical trial (N = 225 at tLM = 4 years); Paquid cohort (N = 1561 at age 80; N = 1240 at age 85)	Mortality risk; (PBC) landmark: 4 years, window: 3 years; (Paquid) landmark: age 80 and 85 years, window: 5 years	10‐fold CV (repeated 50×) on real data; simulation: 250 generated datasets + external validation set (N = 500)	AUC: 0.73–0.87 (PBC); 0.55–0.64 (Paquid). Best Brier Score: 0.076
Cystic fibrosis prognostication via automated ML (2018) [55]	AutoPrognosis (Bayesian optimization + ensembles)	UK CF Registry; 4064 adult CF patients (2012 baseline; FEV1 trajectory 2008–2011)	Mortality/lung transplant risk; landmark: 2012 baseline; window: 3 years	Internal CV via Bayesian optimization (AutoPrognosis framework); no external validation reported	C‐index: 0.89
Dynamic prediction using random survival forests with longitudinal biomarker (2021) [22]	RSF	Heart valve: N = 256 (joineR dataset); kidney transplant: N = 535 (178/33% dnDSA)	Survival/dnDSA risk; (heart valve) landmark: 0.5–3 years, window: 3 years; (kidney) landmark: 6–24 months post‐transplant, window: 1 years	5‐fold CV (100 iterations) for both real‐data applications; simulation: 50/50 train/test split (500 simulated datasets)	AUC: 0.52–0.64 (kidney transplant); 0.58–0.72 (heart valve)
Tumor modeling and survival prediction with Neural‐ODE (2023) [36]	Neural‐ODE	IMPower150 Phase 3 NSCLC trial; 1105 Stage IV NSCLC patients (≥ 2 tumor measurements); 32‐weeks observation window	Overall survival risk; landmark: none; window: extrapolated beyond 32‐weeks input	80/20 random train/test split +5‐fold CV on training set for OS prediction model	C‐index: 0.84 ± 0.02 (CV), 0.82 (test). Tumor RMSE: 9.69, R 2 = 0.88
Lupus nephritis flare prediction using LSTM (2023) [31]	LSTM (RNN)	1694 LN patients, Nanjing Glomerulonephritis Registry (1985–2010); 32,227 longitudinal records; median 4.1‐years follow‐up	Relapse risk; landmark: 3‐visit sliding window; window: variable	80/10/10 random person‐level split (training 25,781/validation 3223/test 3223 data points)	C‐index: 0.897 (IMV‐LSTM); 0.880 (LSTM)
Treatment‐aware survival prediction modeling (2024) [56]	DeepSurv + variant	SEER‐Medicare (1991–2016); 187,167 women with Stage I–III breast cancer, age ≥ 65	All‐cause mortality risk; landmark: none (diagnosis baseline, time‐varying covariates); window: 20 years	5‐fold CV for implementation validation; Bayesian hyperparameter optimization on SEER‐Medicare main analysis	C‐index: 0.60–0.72 (time‐fixed); 0.842–0.996 (time‐varying). IBS: 0.001–0.089
Postoperative complication prediction with multi‐task DNN (2023) [57]	Multi‐task DNN	University of Florida Health; 56,242 patients, 67,481 inpatient surgeries (June 2014–September 2020)	Postoperative complication risk (9 outcomes); landmark: pre/intra/perioperative phases; window: perioperative	Chronological train/test split: development Jun 2014–Nov 2018 (47,188 surgeries); validation Nov 2018–Sep 2020 (20,293 surgeries)	AUROC: 0.62–0.93 (9 outcomes); 0.91–0.93 (primary outcomes)
Dynamic survival model using conditional variational inference (2024) [58]	Conditional VAE + NN	MIMIC‐IV (71,935 ICU stays) + eICU (200,859 encounters); benchmark static datasets: SUPPORT (8873), METABRIC (1904), GBSG (2232), NWTCO (4028)	ICU mortality risk; landmark: 72‐h; window: 24 h	60/20/20 train/validation/test split (all datasets); held‐out test sets; results averaged over 5 random seeds	C_td: 57.9 (MIMIC‐IV), 60.3 (eICU)
Alzheimer's progression prediction via functional survival forests (2019) [47]	Functional RSF	ADNI‐1; 355 MCI patients (same dataset as row 7 Li & Luo, 2019 [47]); simulation: 100 datasets (I = 300)	AD conversion risk; landmark: 0.5, 1, 1.5, 2 years; window: 0.5 and 1 years	10‐fold CV (same paper as row 7); Functional RSF model not explicitly described in paper [18]—may refer to a separate publication	C‐index: 0.80
Cardiovascular event prediction using multivariate longitudinal data (2023) [19]	Explainable ensemble	CARDIA cohort; 3539 participants, 6 exam visits over 15 years (1985–2001); 181 CVD events (5.1%)	CVD event risk; landmark: Y15 exam (2000–01); window: 17 years	5‐fold × 2 repeated CV (stratified)	iAUC: 0.808 (RSF), 0.794 (Dynamic‐DeepHit). C‐index: 0.778/0.767
Dynamic predictive accuracy of joint models and landmarking (2015) [59]	Joint models; Landmarking	Paquid cohort (training, n = 2970); Three‐City cohort (validation, n = 3880; Bordeaux + Montpellier); dementia with competing risk of death	Dementia risk (competing risk: death); landmark: s = 0–4 years (0.5‐years steps); window: 5 years	External cohort validation: model trained on Paquid cohort, evaluated on independent Three‐City cohort	Dynamic AUC: 0.79–0.90 (3C validation); IST uniformly superior to MMSE across all landmark times; Brier Score reported (difference not significant)
Dynamic prediction of multivariate longitudinal data using deep learning (2022) [9]	TransformerJM	ADNI: 1068 MCI patients (331 ad events, mean 3.3‐years follow‐up); NACC: 5261 MCI patients (1516 ad events); simulation: 100 datasets (I = 1000)	AD conversion risk; landmark: 1–4 years; window: 1 year	10‐fold CV (ADNI and NACC); simulation: 70/30 train/test split (100 datasets)	iAUC: 0.890 (ADNI), 0.812 (NACC); iBS: 0.062/0.080

Abbreviations: 3C, Three‐City cohort; ADNI, Alzheimer's Disease Neuroimaging Initiative; CARDIA, Coronary Artery Risk Development in Young Adults Study; CVD, cardiovascular disease; dnDSA, de novo donor‐specific antibodies; eICU, eICU Collaborative Research Database; EUROCARE, European Cancer Registry‐based study; FEV1, forced expiratory volume in 1 s; GBSG, German Breast Cancer Study Group; iBS, integrated Brier score; IST, Isaacs Set Test; MCI, mild cognitive impairment; METABRIC, Molecular Taxonomy of Breast Cancer International Consortium; MIMIC, Medical Information Mart for Intensive Care; MMSE, Mini‐Mental State Examination; NACC, National Alzheimer's Coordinating Center; NWTCO, National Wilms' Tumor Study; OOB, out‐of‐bag; PBC, primary biliary cirrhosis; SEER, Surveillance, Epidemiology, and End Results Program; sPLS‐DR, sparse partial least squares for dimension reduction; SUPPORT, Study to Understand Prognoses and Preferences for Outcomes and Risks of Treatments; VAE, variational autoencoder.

^a Performance metrics are reported as named in source publications. C‐index (concordance index) and AUROC/AUC are the most common metrics; Brier Score, RMSE, and precision/accuracy/recall are reported where C‐index/AUC are unavailable. Multiple metrics are shown where reported, matched to specific ML/DL models.

3.1. Model Architecture

The included studies spanned traditional statistical models, ensemble machine learning methods, and modern deep neural networks. Overall, DL‐based architectures were the most prevalent (roughly 60% of studies), but many analyses also evaluated Cox regression variants and tree‐based ensembles as benchmarks or alternatives. Below, we summarize each model category and its clinical relevance.

3.1.1. Cox‐Based Models

Several studies applied Cox proportional hazards models or extensions thereof as either benchmark methods or comparators. The classical Cox PH model remained a strong baseline—for example, Yang et al. [18] found that a Cox model attained a time‐dependent concordance index of about 0.78 on a colorectal cancer cohort, only marginally below the best ML model's 0.79 (a nonsignificant difference). Cox‐based approaches offer interpretability through hazard ratios, which were highlighted as an advantage in clinical contexts where transparency is valued. Some investigators extended Cox models to better handle longitudinal data, using time‐dependent covariates or non‐proportional hazards formulations [47, 60]. Others employed penalized Cox techniques (LASSO, elastic‐net) to select prognostic features in high‐dimensional settings [19, 42]. These regularized Cox models were particularly useful in genomic or multiomic studies, yielding compact risk scores. In general, Cox‐based models demonstrated competitive discrimination in many datasets and serve as a baseline for dynamic methods. However, their reliance on proportional hazard assumptions and linear covariate effects can limit performance when relationships are highly nonlinear or time‐varying, conditions where ML/DL models tend to excel.

3.1.2. Tree‐Based Models

Ensemble tree methods (especially random survival forests and gradient boosting machines) were frequently used for dynamic prediction. These nonparametric models automatically capture interactions and nonlinear effects, often improving accuracy over Cox models in heterogeneous data. For instance, Sato et al. [60] reported a time‐dependent AUC of 0.785 for an RSF predicting prostate cancer recurrence, outperforming a Cox model's 0.704 and even clinical nomograms. Similarly, in Yang et al.'s [27] colorectal cancer study, the RSF achieved slightly higher C‐index (0.786) than Cox (0.781). Tree‐based approaches also offered interpretability via variable importance measures; several studies applied SHAP (Shapley Additive Explanations) to RSF or boosting models to identify key prognostic factors. Both vanilla and variant forms were explored: Wongvibulsin et al. [52] introduced an ensemble (RF‐SLAM) tailored for simultaneous survival and longitudinal outcome modeling, while others used functional survival forests to directly incorporate entire biomarker trajectories [21]. Gradient boosting machines (e.g., XGBoost) were popular and showed performance on par with RSF in many cases: for example, Wang et al. [20] reported a boosted‐tree (XGBoost/GBM‐family) survival model achieving discrimination comparable to RSF for predicting all‐cause mortality in metabolic dysfunction‐associated fatty liver disease, and Zheng et al. [26] reported similar boosted‐tree performance for outcome prediction in pediatric Hodgkin and non‐Hodgkin lymphoma. Moncada Torres [61] demonstrated that an explainable boosted‐survival model could even outperform Cox regression for breast cancer prognosis. The clinical appeal of tree‐based models lies in their robustness to complex data (handling missing values, nonlinear effects) and the ease of generating decision rules or risk scores. In summary, tree‐based methods often improved risk discrimination, particularly when patient populations or risk factors were diverse, while maintaining a degree of interpretability through variable importance or surrogates.

3.1.3. Neural Network Models

Deep learning approaches constituted the largest category of dynamic prediction models in our review. These ranged from extensions of Cox models with neural networks (e.g., Deep‐Surv) to completely novel architectures like recurrent neural networks and transformers. Many studies included DeepSurv or similar deep Cox nets as benchmarks, finding them competitive with tree ensembles (often reporting C‐index values around 0.78–0.80 in validation). Fully sequential models were also prominent: Long Short‐Term Memory (LSTM) and other RNN variants were used to ingest patients' time‐series data (such as vital signs or lab results) and update risk continuously. In a critical care setting, Shashikumar [54] developed an interpretable RNN survival model (DeepAISE) that predicted sepsis onset 4 h in advance with an AUC of 0.90, substantially earlier and more accurately than conventional warning scores. Similarly, in chronic disease, RNN‐based models could leverage longitudinal patterns of biomarkers to anticipate events like renal flares or heart failure decompensation [7, 31]. Transformer‐based architecture has recently emerged for survival analysis of longitudinal data; these use self‐attention to capture long‐range dependencies. For example, a Transformer model with an embedded joint learning scheme (TransformerLSR) was proposed to jointly model longitudinal measurements, survival outcome, and recurrent events, thereby capturing complex interdependencies in patient history [37]. While such comprehensive models are rare, they illustrate the flexibility of deep learning—TransformerLSR can learn how intermediate events (e.g., hospitalizations) influence survival risk, which is clinically insightful. Other innovations included multi‐task networks (predicting multiple outcomes simultaneously) and Neural‐ODE‐based continuous‐time models that integrate mechanistic dynamics [36]. Notably, deep models achieved some of the highest performance metrics reported: for instance, in a large prostate cancer cohort, a DL model attained a C‐index up to 0.85 for 5‐ and 10‐year survival prediction [53]. However, authors also cautioned that the data‐driven power of DL comes with requirements for large sample size and careful validation. In one study, ML/DL models did not significantly outperform Cox due to limited sample size, suggesting that their potential advantages manifest primarily in “big data” settings [15]. In practice, the advantage of neural networks lies in their ability to model complex, nonlinear temporal relationships (e.g., interactions between treatments and biomarkers over time), at the cost of interpretability. Efforts to mitigate the black‐box nature (attention weights, saliency maps, etc.) were evident in several studies leveraging DL, aiming to ensure clinical relevance of the learned risk features.

3.2. Temporal Design

A key aspect of dynamic prediction is how time is handled. The included studies varied in whether they treated time in discrete steps or continuously, whether and how often they updated predictions (dynamic vs. static models), how longitudinal covariates were incorporated, and the choice of observation and prediction windows. We analyze each of these temporal design choices below.

3.2.1. Discrete‐Time Models

Many ML/DL models formulated the survival problem in discrete time, dividing follow‐up into intervals and predicting event occurrence probabilities in each interval. This approach was especially common among neural network models (e.g., DeepHit, dynamic‐deep neural nets), as it converts survival analysis into a series of classification sub‐tasks. For example, Lee et al. [35] introduced Dynamic‐DeepHit, which outputs a probability distribution over discrete time bins for each outcome/risk. Discrete‐time formulations offer flexibility in handling censoring and competing risks—one can simply compute loss terms at each time bin where an individual is at risk [35, 59, 62]. Indeed, over half of the deep learning models in our review employed a discrete‐time hazard output (see Table 2). A practical benefit is that arbitrary predictive horizons—i.e., the ability to produce a risk estimate for any user‐specified follow‐up time without retraining the model, simply by selecting the corresponding time bin(s) (discussed further in Section 3.2.5)—can be accommodated by choosing the appropriate time‐bin span (e.g., monthly risk predictions up to 1 year, as demonstrated in some implementations of DeepHit). However, the predictions can be sensitive to the chosen interval granularity. Coarser intervals reduce temporal resolution of the risk estimates, while very fine intervals increase computational load and may require larger data to estimate reliably. Suresh et al. [4] provide a tutorial on discrete‐time survival models, emphasizing these trade‐offs. Overall, discrete‐time dynamic models have proven effective and were the dominant paradigm for DL‐based survival prediction in the literature.

3.2.2. Continuous Time Models

Other methods operated in continuous time, modeling risk without predefined time bins. Classical Cox models are inherently continuous‐time, and newer techniques maintain this trait. Neural network adaptations of Cox (such as DeepSurv and Cox‐Time) optimize partial likelihood across continuous event times, avoiding any need for discretization [58, 63]. Likewise, recent Neural‐ODE survival models learn a smooth hazard function over time [36]. These continuous‐time approaches avoid the bias that can be introduced by binning time and naturally handle irregularly spaced data points. For instance, Laurie and Lu [36] employed a neural ODE to integrate longitudinal tumor measurements and predict overall survival; this physics‐inspired continuous model was able to extrapolate tumor growth trajectories and achieved higher C‐indices than traditional tumor growth curve modeling. Continuous‐time models align well with scenarios like pharmacotherapy monitoring, where events (and drug exposures) occur in real clock time, and one may wish to update risk at any moment. While these models can be more complex to implement (sometimes requiring specialized solvers or loss functions), they offer the advantage of fine temporal precision. The ability to make “instantaneous” risk predictions is valuable for integrating real‐time sensor data or continuous drug concentration profiles. On the other hand, purely continuous models may need strong assumptions (e.g., smoothness of hazard over time) and can struggle if data are extremely sparse over time. Hybrid approaches also exist (e.g., combining continuous underlying hazard with discrete monitoring times via pseudo‐observations [64]) but were less commonly seen.

3.2.3. Dynamic Versus Static Prediction

A defining feature of “dynamic” models is that they update the risk prediction whenever new patient data becomes available. In contrast, static models use only baseline information (e.g., at time of diagnosis) to predict an outcome at a future time and do not change that prediction. This dynamic/static distinction corresponds to the “Dynamic” column of Table 2, indicating for each algorithm family whether implementations updated predictions as new data arrived (“Yes”) or used a fixed baseline (“No”). Most studies we reviewed aimed for dynamic updating: as new labs, imaging, or visits occurred, the model could revise the patient's risk. Continuous models like RNNs or Cox models with time‐varying covariates inherently do this—each new observation shifts the risk estimate. An alternative strategy observed in some papers is landmarking, a semi‐dynamic approach. Landmark models specify fixed time points (landmarks) and use data up to each landmark to produce a fresh survival prediction beyond that point. Several works adopted landmark analysis to balance complexity and practicality [5, 6, 22]. For example, Devaux et al. [5] used a landmark Superlearner (combining Cox and RSF) at yearly intervals to dynamically predict 1–3‐year risk of an event as patients' biomarker histories evolved. This yielded more accurate individual predictions compared to using only baseline values [19, 20, 47]. True dynamic models can utilize all interim data more fluidly, but the landmark approach is often easier to implement with traditional models and still captures some benefits of updating. In contrast, purely static models (nondynamic) were rare among included studies—partly because our inclusion required longitudinal data. Nonetheless, a few comparisons emphasized the value of dynamic updating: in Dai et al. [53], including longitudinal covariates like PSA values enables more accurate and clinically relevant short‐term risk predictions by capturing dynamic disease progression over time. Generally, dynamic models showed equal or better discrimination than static ones, especially for longer‐term horizons where baseline risk profiles can be overtaken by intervening health changes.

3.2.4. Use of Longitudinal Covariates

How models utilized longitudinal covariate information varied widely. Simpler dynamic methods often reduced longitudinal data to a single time‐dependent covariate per variable—for instance, a time‐dependent Cox model might input the most recent lab value or cumulative dose at each time [19]. This approach captures the latest status but may ignore trajectories. More advanced models attempted to leverage the entire history of biomarkers or clinical measurements. Some statistical approaches, like joint modeling and functional data analysis, explicitly incorporate whole longitudinal trajectories. Joint modeling of longitudinal and survival data has been widely used in oncology and immunology trials to link biomarker trajectories with clinical outcomes. Internal applications within Sanofi and other pharmaceutical settings have shown that joint models can provide early decision‐making in oncology trials [48]. Pickett et al. [22] introduced a random survival forest that uses each subject's biomarker trajectory as a functional covariate, thereby accounting for the shape of the biomarker curve over time. Similarly, a functional survival forest approach successfully captured multivariate longitudinal biomarker trajectories in Alzheimer's disease [21]. In the ML realm, multi‐step features (slopes, variability) were sometimes engineered to summarize longitudinal patterns for tree models [19]. Deep learning models, on the other hand, typically ingest the raw sequence of longitudinal inputs: RNN‐based architectures accumulate information from all previous time points, and attention‐based models can learn which past observations are most relevant to the prediction. For example, an RNN may implicitly learn that a downward trend in lung function over several visits portends higher risk, even if absolute values are still normal. Similarly, transformer models can attend specific past events (e.g., a spike in inflammatory markers) that influence current risk [37]. A few studies also pursued multitask training, where the model simultaneously predicts future values of a longitudinal biomarker and the survival outcome [9, 57]. This way, the model is encouraged to extract prognostically relevant features of the longitudinal data. Generally, the more of the longitudinal signal a model can exploit, the better the predictive performance—although this benefit often comes at the cost of greater model complexity and larger data requirements.

3.2.5. Observation Windows and Prediction Horizons

Dynamic prediction models are also characterized by the length of past data used (observation window) and how far into the future they predict (prediction horizon). In practice, many researchers evaluated their models at specific clinically relevant horizons. Common choices were 30‐ or 90‐day mortality for acute conditions, and 1‐, 3‐, or 5‐year survival for chronic diseases or cancers (see Table 3 for more details). Zhu et al. [65], for example, built an ML model to predict 30‐day mortality in hemophagocytic lymph histiocytosis, an acute hyperinflammatory syndrome. In contrast, Yuan et al. [66] developed a deep survival model for dementia risk with a much longer horizon (prediction of incident dementia over decades). The observation window is the span of historical data given to the model at the time of prediction. In ICU settings with high frequency data, observation windows might be as short as the last 6–12 h of vitals to predict the next 24 h [51]. Shashikumar [54] used approximately 8 h of ICU data to inform their 4‐h ahead sepsis predictions. By contrast, in chronic disease management, an observation window could encompass years of outpatient data; for example, a model might use 5 years of longitudinal lab results to predict risk of an event in the next year. Some studies explicitly experimented with different observation window lengths or update frequencies. Zheng et al. [32] developed a real‐time ICU risk prediction model that handled irregularly sampled labs and vitals, using all available data up to the present to update risk continuously. Generally, tailoring the observation window to the clinical context is important: it should be long enough to capture relevant history (e.g., baseline condition and recent changes) but not so long as to include obsolete information. Prediction horizon is similarly context‐dependent—short horizons for acute decompensation (to allow timely intervention), and longer horizons for guiding long‐term prognostication. The studies in our review illustrate this balance: dynamic models have been successfully applied from the sub‐hourly scale in sepsis prediction to the multi‐year scale in oncology outcomes, by selecting clinically appropriate windows and horizons.

3.3. Drug Exposure and PK/PD Integration in Dynamic Prediction

In many therapeutic areas, instantaneous clinical risk (e.g., death, progression, flare, hospitalization) is mechanistically linked to time‐varying drug exposure and downstream pharmacodynamic (PD) effects [64, 67, 68, 69]. Ting et al. [70] compare a sequential and joint model of tumor kinetics and survival after Durvalumab treatment in patients with metastatic urothelial carcinoma. Marchand et al. [71] study tumor growth and overall survival modeling to support decision making in phase Ib/II trials. Han et al. [72] integrate modeling of biomarkers, survival and safety in clinical oncology drug development.

When models ignore exposure heterogeneity, apparent treatment effects are attenuated and the timing of benefit is obscured; by contrast, integrating pharmacokinetic/pharmacodynamic (PK/PD) information allows individualized risk to be updated as dosing changes, concentrations fluctuate, and biomarkers respond, aligning predictions with real decision points in titration, therapeutic drug monitoring (TDM), and benefit–risk assessment [67].

For example, Poon et al. [16] demonstrated that correctly specified Cox models can recover exposure–response relationships in large‐molecule oncology drugs, while mis‐specifying exposure functions or ignoring confounding distorts the apparent hazard. In practice, longitudinal exposure histories can be constructed either directly from measurements (e.g., trough and peak levels, time‐above‐threshold, rolling or cycle AUC with clinically meaningful lags) or via model‐informed exposures obtained from a population PK model that yields individual C(t) or AUC(t) trajectories and, when relevant, effect‐site signals (e.g., E(t) or inhibition indices) [8, 67].

Because many exposure–response pathways are mediated by longitudinal PD biomarkers, it is beneficial to include both PK (e.g., time‐dependent concentration C(t), AUC[t]) and PD covariates in survival modeling. Mechanistic reviews emphasize that PK–PD integration enables prediction of clinical outcomes by linking dose → concentration → effect → response [73]. Empirically, Hansson et al. [74] showed in sunitinib‐treated GIST patients that when model‐predicted PD biomarkers (sVEGFR‐3 REL and tumor‐size change) were included together with exposure (AUC) in a Weibull time‐to‐event model, the PD covariates supplanted AUC as dominant predictors of overall survival, improving model fit (ΔOFV = 51.6) and calibration. This demonstrates that PK exposure informs event risk primarily through its downstream PD effects and that using both PK and PD trajectories yields the most accurate survival predictions. Similar results have been observed in other oncology PK/PD frameworks, where PD variables (e.g., tumor‐size dynamics, receptor occupancy, cytokine changes) captured residual variability and mediated exposure–response effects [73]. However, when both PK exposure metrics and PD/biomarker covariates are entered as joint predictors, confounding must be considered: biomarkers such as serum albumin can be simultaneously prognostic for survival and a determinant of drug clearance [75, 76], so failing to account for this shared dependency may bias the estimated exposure—response relationship and overstate or mask the contribution of exposure to risk [77].

Figure 2 illustrates a proposed framework for integrating AI models into survival analysis in the setting of oncology, in combining with pharmacokinetic (PK) modeling based on drug concentration–time data. A PK model estimates key parameters such as clearance (CL) and volume of distribution (V), along with their variability across individuals. Patient‐specific PK profiles can be derived, and baseline covariates—such as gender and body weight—can be incorporated to account for inter‐patient variability. A well‐developed population PK model provides insight into how the body processes the drug. When these models are coupled with machine learning and deep learning techniques in survival analysis, they can significantly enhance predictive performance for clinical endpoints, including adverse effects, tumor response, and overall survival.

FIGURE 2.

A model‐based framework integrating clinical oncology drug development with machine learning (ML) and deep learning (DL) techniques enables more informed prediction of clinical outcomes. Starting from the development of a population pharmacokinetic (PK) model, derived PK metrics—such as dose intensity, area under the curve (AUC), and concentration‐time profiles—can be incorporated into pharmacokinetic‐pharmacodynamic (PKPD) models to evaluate various responses, including adverse events and tumor dynamics (TSR: Tumor size ratio and Tumor (t): Tumor time course). These PKPD models also serve as tools for predicting survival outcomes (OS: Overall survival and PFS: Progression‐free survival) and guiding dose and regimen adjustments.

Although only a limited number of studies have applied machine‐learning (ML) and deep‐learning (DL) models to detect or predict PK/PD‐related clinical endpoints, these approaches show clear advantages over traditional statistical methods in handling complex, nonlinear, and time‐dependent exposure–response relationships. Several modeling families support PK/PD‐aware dynamic prediction, each with clinical advantages. Cox models with time‐dependent covariates accommodate C(t), AUC(t), dose intensity, and lagged exposure terms, and retain interpretability for clinicians. Landmarking offers a pragmatic alternative when updates occur at scheduled cycle boundaries. Tree‐based survival methods (RSF, GBM/XGBoost) flexibly learn thresholds and high‐order interactions between exposure and comorbidities and provide variable‐importance profiles that highlight clinically interpretable exposure drivers [8]. Neural survival models extend this flexibility: DeepSurv/Cox‐time optimize continuous‐time hazards with exposure inputs; discrete‐time hazard networks (e.g., DeepHit variants) output interval risks as functions of exposure histories; multitask networks jointly forecast biomarker trajectories and hazards; and Neural‐ODE survival links mechanistic disease dynamics to event risk [9, 36]. Deep joint models—including transformer architectures—capitalize on long‐range dependencies and can simultaneously represent longitudinal measurements, survival, and even recurrent events, thereby learning how intermediate states (e.g., hospitalizations or toxicity) shape subsequent risk [37].

3.4. Regulatory Perspective

Dynamic survival models based on ML/DL methods are beginning to intersect with the broader framework of Model‐Informed Drug Development (MIDD). Their potential application in dose selection, trial adaptation, or patient stratification naturally raises questions of credibility, transparency, and regulatory acceptance. Two recent draft guidances are particularly relevant. The ICH M15 guideline on general principles for Model‐Informed Drug Development, finalized in 2026, proposes a harmonized framework for evaluating MIDD evidence, emphasizing clarity of the question of interest, context of use and model risk/impact, and the need for transparent documentation [78]. Similarly, the FDA's draft guidance on the use of AI to support regulatory decision‐making outlines a risk‐based credibility framework, stressing that the extent of verification, validation, and external evaluation should be proportional to the regulatory influence of the model [79].

Mapped onto the use cases reviewed here, the ICH M15 general principles for model‐informed drug development [78] and the FDA's draft guidance on AI for regulatory decision‐making [79] both provide a risk‐based credibility assessment structure directly applicable to dynamic prediction models. Central to both frameworks is the concept of context of use (COU): whether a model's output, such as a 1‐year mortality probability updated at each clinic visit, informs an exploratory analysis or a primary decision such as dose selection or trial enrichment, and what an incorrect prediction would mean for patient safety or program risk. Applied to the dynamic models in Table 2, a higher‐impact COU (e.g., a Dynamic‐DeepHit‐type model used to trigger dose modification) would warrant correspondingly more extensive credibility evidence in terms of broader external validation and sensitivity analyses than a lower‐impact COU (e.g., an exploratory risk‐stratification tool used for hypothesis generation).

A related consideration is that the ICH M15 fit‐for‐purpose principle applies differently across the model classes reviewed in Sections 3.1, 3.2. Joint models and landmark Cox‐based approaches are built on explicit, parametric assumptions about the longitudinal and survival sub‐models, which facilitates credibility assessment because each component can be checked against domain knowledge (e.g., plausibility of estimated hazard ratios for PK/PD covariates). In contrast, black‐box deep learning architectures (e.g., RNN, transformer, or Neural‐ODE survival models) achieve strong discrimination (Table 2), but their internal representations are not directly interpretable, so establishing fit‐for‐purpose credibility relies more heavily on empirical performance evidence: external validation in independent cohorts, calibration assessment, and explainability analyses (e.g., attention weights or SHAP‐type variable‐importance measures) that approximate, but do not replace, mechanistic justification. Under a risk‐based framework, black‐box DL models intended for higher‐impact contexts of use should therefore be held to a higher evidentiary bar for external validation and explainability than statistically grounded joint or landmark models used in similarly impactful contexts.

Beyond ICH M15 and the FDA's AI guidance [79], dynamic ML/DL survival models deployed in clinical decision support may also fall within broader regulatory frameworks for AI‐enabled technologies. The IMDRF's SaMD framework [80] and FDA's Good Machine Learning Practice (GMLP) principles [81] both emphasize lifecycle management—including data quality, model monitoring, and performance‐drift detection—considerations especially relevant for dynamic models periodically retrained on accumulating longitudinal data. In the EU, the AI Act [82] introduces risk‐tiered obligations, and AI‐based clinical prediction tools informing treatment decisions could be classified as high‐risk, triggering requirements for risk management, technical documentation, and human oversight. Sponsors developing dynamic prediction models for MIDD or clinical decision support should therefore anticipate that these broader AI‐governance requirements may apply alongside the MIDD‐specific credibility expectations of ICH M15.

A recurring expectation across these frameworks is external validation, transportability, and subgroup‐level performance characterization. As discussed in Section 4.1, however, most studies in our review reported only internal validation—cross‐validation (CV) or a single train/test split within one institution—with limited evidence of performance across external cohorts, different healthcare systems, or clinically important subgroups. From a regulatory perspective, demonstrating that a dynamic prediction model maintains adequate discrimination and calibration when transported to a new population, and that performance does not vary substantially across subgroups (e.g., by age, sex, race/ethnicity, disease severity, or comorbidity burden), is likely to be a prerequisite for any context of use with meaningful patient or trial impact [78, 79].

Both FDA's AI guidance and ICH M15 guideline signal a strong regulatory interest in how advanced modeling, including dynamic ML/DL survival methods, will be evaluated. For pharmacometricians, this guidance highlights that fit‐for‐purpose evaluation, risk awareness, and transparent reporting will be critical for the future acceptance of dynamic prediction approaches.

This is especially pertinent for models that integrate PK/PD features, where predictions must align with clinically meaningful decision points such as dose titration or therapeutic drug monitoring. In such cases, regulatory credibility depends not only on predictive accuracy but also on clear documentation of exposure derivation, external validation, and explicit characterization of uncertainty. Taken together, ICH M15 and FDA's AI guidance document provide emerging principles that can help orient the design, evaluation, and communication of dynamic survival models. Adopting these principles early can bridge methodological innovation with regulatory expectations, preparing the field for a future where dynamic prediction contributes directly to decision‐making in drug development.

4. Discussion

Dynamic ML/DL models are being applied across a wide range of clinical contexts. In acute care, recurrent and convolutional networks ingest high frequency vital‐sign and laboratory streams to forecast sepsis or mortality hours ahead; the interpretable DeepAISE model, for example, predicted sepsis 4 h in advance with an AUROC around 0.90 [54]. Other real‐time models in ICU cohorts achieve AUROC or C‐index values exceeding 0.85 using irregularly sampled data streams [32, 37]. In chronic disease and oncology, dynamic survival models leverage years of longitudinal data to predict multi‐year survival or progression; examples include deep survival machines for prostate cancer [53], functional random survival forests for Alzheimer's disease [47], and transformer joint models that learn from multivariate biomarker histories [9]. Landmarking and super‐learner ensembles offer a pragmatic compromise between fully dynamic updating and computational simplicity, yielding competitive discrimination for 2‐ to 5‐year horizons [5, 6].

Across the studies reviewed, model architecture and temporal design were closely linked: Cox‐based and tree‐based models operated predominantly in continuous time with fixed, non‐updating predictions, while deep learning architectures—particularly recurrent, transformer‐based, Neural‐ODE, and Dynamic‐DeepHit‐type models—were more often designed to incorporate longitudinal measurements and update predictions dynamically as new data arrived. This architectural shift toward deep learning has thus been accompanied by a parallel shift toward dynamic, repeatedly updated risk prediction, with landmarking remaining the principal strategy by which traditional Cox‐based models approximate dynamic updating without requiring a time‐varying architecture.

Despite promising performance, current evidence highlights several gaps. Temporal design choices such as observation window length, prediction horizon and update frequency vary widely, and many studies report only discrimination metrics (e.g., C‐index) without calibration or clinical‐utility analyses. As shown in Table 3, the reviewed studies employ a broad range of evaluation metrics (AUC, C‐index, RMSE, and study‐specific statistics), reflecting the methodological diversity of this emerging field; a narrative synthesis was therefore adopted to enable meaningful comparison across heterogeneous contexts. Consistent with this, calibration is reported only sporadically and could not be summarized systematically. External validation on independent cohorts is rare. Importantly, only a handful of models explicitly incorporate PK/PD information: most use time‐varying clinical covariates, while few studies have leveraged population‐PK derived covariates, and only one multitask network has jointly modeled drug concentrations with hazards [7, 8]. Integrating dose intensity, concentration–time profiles and lagged exposures into ML models can align risk predictions with titration and therapeutic‐drug monitoring decisions and improve discrimination [7, 16, 73], but methodological demonstrations remain sparse.

A further consideration not addressed above is the handling of censoring. Most models reviewed assume censoring is non‐informative—that is, independent of the event process conditional on covariates—an assumption inherited from the Cox partial‐likelihood framework and extended into many ML/DL adaptations, including discrete‐time hazard losses and DeepSurv‐type partial‐likelihood losses. This assumption may be violated in practice when, for example, patients are lost to follow‐up due to worsening health or treatment discontinuation linked to perceived futility. A subset of included studies applied IPCW for performance metric estimation [5, 6, 59], and joint models [9, 47] partially address informative dropout by simultaneously modeling longitudinal and event processes; however, none explicitly modeled or tested for informative censoring during model training. This remains an important methodological gap, particularly for models intended to support regulatory decision‐making, where biased risk estimates could have direct clinical consequences.

From a regulatory standpoint, guidance such as the now‐finalized ICH M15 and the FDA's draft considerations on AI emphasize principles rather than prescriptive workflows, highlighting context‐of‐use definition, risk–impact assessment, and transparent reporting as essential elements of credibility. Interpreting this guidance as emerging principles can help orient the development and communication of dynamic survival models, ensuring they evolve in a way that is both scientifically rigorous and aligned with future expectations for regulatory review.

Moving forward, standardized reporting of temporal parameters, calibration alongside discrimination, and external validation is essential. To facilitate clinical adoption, dynamic models should incorporate interpretability tools, such as attention mechanisms or variable‐importance measures, and align with evolving regulatory frameworks. Hybrid approaches that embed mechanistic PK/PD or disease models within ML architectures may offer the dual advantages of interpretability and predictive accuracy [8, 9]. Establishing shared benchmarks and longitudinal datasets will further enable fair comparison across methods and accelerate translation. In conclusion, integrating pharmacometrics with AI/ML‐enabled dynamic prediction offers a path to more precise dosing, refined patient stratification, and adaptive trial design, while aligning innovation with emerging regulatory expectations to maximize clinical impact.

4.1. Limitations

This review should be interpreted in light of several limitations. First, the search was restricted to PubMed. While this approach is consistent with the focused scope of this review and with many rapid or scoping reviews, it may have missed relevant studies indexed exclusively in other databases, such as Embase, Scopus, or Web of Science, or studies reported in the gray literature. Second, we did not perform a formal risk‐of‐bias or methodological quality assessment of the included studies, such as PROBAST for prediction‐model studies. Therefore, the reported model performance estimates should be interpreted with caution, recognizing that methodological quality and validation rigor varied across studies. Third, because the included studies differed substantially in study populations, endpoints, prediction horizons, validation designs, and reported performance metrics, a meta‐analysis or pooled quantitative synthesis was not feasible. Instead, this review provides a qualitative and descriptive synthesis of reported methods, model types, validation strategies, and performance patterns across the literature. Future updates could build on this work by incorporating a multi‐database search strategy, formal quality appraisal, and, where sufficient methodological consistency exists, meta‐analytic synthesis of comparable performance metrics.

Author Contributions

Wrote manuscript: Xinyi Liu, Malidi Ahamadi. Designed research: Malidi Ahamadi, Shuai Shao, Dorothee Semiond, Xinyi Liu. Performed research: Xinyi Liu, Malidi Ahamadi, Shuai Shao. Analyzed data: Xinyi Liu.

Funding

The authors have nothing to report.

Conflicts of Interest

Xinyi Liu, Shuai Shao, Dorothee Semiond, and Malidi Ahamadi are employees of Sanofi and may hold stock and/or stock options in Sanofi.

Liu X., Shao S., Semiond D., and Ahamadi M., “Dynamic Prediction of Clinical Endpoints in Survival Analysis With Machine and Deep Learning: A Systematic Review and Regulatory Perspective,” CPT: Pharmacometrics & Systems Pharmacology 15, no. 10 (2026): e70330, 10.1002/psp4.70330.