Joint models in both directions and at scale

jmBIG and JMbdirect for dynamic prediction

Repeated biomarkers and the events they predict, modelled together, when events also change the biomarker and the cohort runs to a million patients

Atanu Bhattacharjee

Senior Lecturer and Academic Statistician, School of Medicine, University of Dundee

Royal Statistical Society

How we arrived at the joint model: three routes in

Three separate problems of the late 1980s and early 1990s pointed to one answer.

Survival analysis Cox regression needs the marker's true value at every event time; CD4 was measured with error and only at clinic visits. Gave: a hazard written on the true marker mᵢ(t), not on Yᵢ(t) Prentice 1982 (attenuation) · Self & Pawitan 1992 (marker and onset) Tsiatis, DeGruttola & Wulfsohn 1995 (two-stage, CD4 and survival) Longitudinal analysis Patients dropped out as they became sicker, so naive estimates of decline were biased. Gave: random effects shared between the marker and the dropout process Wu & Carroll 1988 · Wu & Bailey 1989 · Diggle & Kenward 1994 · Follmann & Wu 1995 The clinical question Could CD4 decline stand in for progression to AIDS or death in AIDS clinical trials? Gave: the motivating data, and the need to model the marker's natural history DeGruttola, Lange & Dafni 1991 · Pawitan & Self 1993 Joint model mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ, Yᵢ(t) = mᵢ(t) + εᵢ(t) hᵢ(t) = h₀(t) exp{ γᵀwᵢ + α mᵢ(t) } Faucett & Thomas 1996 · Wulfsohn & Tsiatis 1997 Henderson, Diggle & Dobson 2000

How joint modelling has progressed, 1994 to now

19942001200920202026

Formulation

1994 DeGruttola & Tu: CD4 path and survival

1996 Faucett & Thomas: Bayesian fit by Gibbs sampling

1997 Wulfsohn & Tsiatis: joint likelihood by EM

2000 Henderson, Diggle & Dobson: latent Gaussian processes

Robust estimation, richer models

2001 Tsiatis & Davidian: conditional score, no normality needed

2002 Song, Davidian & Tsiatis: semiparametric random effects

2005–07 several markers; competing risks

2009 Rizopoulos, Verbeke & Lesaffre: Laplace for many random effects

Prediction and software

2009 Proust-Lima & Taylor: dynamic prostate cancer tool from PSA

2010–18 JM, JMbayes, joineRML make fitting routine

2011 Rizopoulos: dynamic predictions and their accuracy

2013–19 web calculators; personalised biopsy schedules

Scale and structure

2020 Dynamic-DeepHit: a deep-learning alternative

2022 FastJM: linear-time fitting for biobanks

2024 jmBIG: partitioned fitting, million-patient cohorts

2026 JMbdirect: events feed back, links bend

Each phase answered the limit of the one before: first how to fit the model, then how to fit it robustly and for more data types, then how to use it for each patient, and now how to fit it at health-record scale and in both directions.

Sources for each entry are listed in the references. Several markers: Brown et al. 2005; Chi & Ibrahim 2006. Competing risks: Elashoff et al. 2007. Web calculator: Taylor et al. 2013. Biopsy schedules: Tomer et al. 2019.

The last fifteen years, 2011–2026: what moved

Four shifts: predictions for each patient, faster fitting, richer models tested head to head, and events that change the marker.

2011 Dynamic predictions patient-specific survival updated at each visit Rizopoulos 2012 The textbook joint models with applications in R Rizopoulos 2013 Web calculator; stjm real-time PSA predictions; joint models in Stata Taylor et al.; Crowther et al. 2014 Model averaging Bayesian weights by patient and by time Rizopoulos et al. 2015 DIA working group Bayesian joint models: current methods, issues Gould et al. 2016 JMbayes Bayesian fitting with dynamic predictions Rizopoulos 2017 JM vs landmarking head to head on calibration, discrimination Rizopoulos et al. 2018 stan_jm; joineRML Stan-based and several- marker fitting in R Brilleman et al.; Hickey et al. 2019 Biopsy schedules predictions set the next clinical test Tomer et al. 2020 Dynamic-DeepHit recurrent network over longitudinal records Lee, Yoon & van der Schaar 2021 Landmarking 2.0 bridging landmarking and joint models Putter & van Houwelingen 2022 FastJM; TransformerJM linear-time biobank fits; attention-based prediction Li et al.; Lin & Luo 2023 INLA joint models fast Bayesian fits, many markers, competing risks Rustand et al. 2024 jmBIG partitioned fitting, a million patients < 1 hour Bhattacharjee et al. 2025 2026 JMbdirect events feed back, links bend Bhattacharjee JMbayes2 0.6 mixed-type markers, competing risks, multi-state Rizopoulos et al.

● software and guidance

● prediction in practice

● speed and scale

● deep learning

● this work

Full references for each entry are on the reference slides.

One patient, three conditions that act on each other

A woman with type 2 diabetes is seen every few months. Her HbA1c and eGFR are measured at each visit. In year four she has a cardiovascular event, and both markers change course.

next two years: second event? HbA1c eGFR cardiovascular event today 0248 years

Prediction

What is her risk over the next two years, and does it change after today's visit?

Direction

The first event moved both of her marker paths. Should the model allow that?

Shape

Does each unit of HbA1c add the same risk, or is there a threshold?

Scale

She is one of a million such patients in routine health records.

Illustrative patient; values are schematic. The diabetes, cardiovascular and kidney feedback loop is the motivating setting in Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

What we observe for each patient

Two linked records: many rows of biomarker values per patient, and one row giving the follow-up time and whether the event happened.

Longitudinal record idtimemarker 1t₁₁y₁₁ 1t₁₂y₁₂ 1t₁₃y₁₃ 2t₂₁y₂₁ 2t₂₂y₂₂ ⋮⋮⋮ Survival record idtimeevent 1T₁δ₁2T₂δ₂ Patient i all marker rows + (Tᵢ, δᵢ) linked by id

Longitudinal outcome

Yᵢⱼ, the marker for patient i at visit time tᵢⱼ; visits are irregular and differ in number between patients.

Time-to-event outcome

Tᵢ = min(Tᵢ*, Cᵢ), the earlier of event time Tᵢ* and censoring time Cᵢ; δᵢ = 1 if the event was observed.

Internal marker

Measured on the patient and only while the patient is event-free, so its path carries information about the event.

Time-to-event data: survival, hazard and Cox

Event times are right-censored. Survival is described by the survival function or the hazard, estimated by Kaplan–Meier, and related to covariates by Cox regression.

time since study entry end of study eventcensored

Survival function

Probability of remaining event-free beyond t.

S(t) = Pr(T* > t)

Hazard

Instantaneous event rate among those still at risk at t.

h(t) = limΔ→0 Pr(t ≤ T* < t+Δ | T* ≥ t)/Δ

Kaplan–Meier

nⱼ at risk and dⱼ events at event time tⱼ.

Ŝ(t) = ∏tⱼ ≤ t (1 − dⱼ / nⱼ)

Cox proportional hazards

Covariates act multiplicatively on an unspecified baseline hazard h₀(t); exp(β) is the hazard ratio.

h(t | x) = h₀(t) exp(xᵀβ)

Censoring is assumed non-informative: given covariates, being censored carries no information about the event. Kaplan EL, Meier P. JASA 1958; 53: 457–481. Cox DR. J R Stat Soc B 1972; 34: 187–220.

Longitudinal data: mixed models and dropout

Repeated measurements are correlated within a patient, irregularly timed and measured with error. The linear mixed model separates the population path from each patient's own deviation.

time marker population average Xᵀβ patient paths Xᵀβ + Zᵀbᵢ

Yᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ + εᵢ(t),  bᵢ ~ N(0, D),  εᵢ(t) ~ N(0, σ²)

Fixed and random effects

β: population path. bᵢ: patient's own intercept and slope deviation; D its covariance, which induces within-patient correlation. σ²: measurement-error variance.

Missing at random (MAR)

Dropout depends only on observed values; likelihood inference from the mixed model remains valid.

Missing not at random (MNAR)

Dropout depends on unobserved values, typically because the event itself ends follow-up. Selection, pattern-mixture or shared-parameter (joint) models are then required.

Laird NM, Ware JH. Biometrics 1982; 38: 963–974. Rubin DB. Biometrika 1976; 63: 581–592. Diggle P, Kenward MG. Appl Stat 1994; 43: 49–93.

The time-varying Cox model and its three assumptions

The extended Cox model uses the covariate value at each event time. That is valid for an external covariate; a biomarker is internal, error-prone and seen only at visits.

time since baseline biomarker event: no further measurements true path mᵢ(t) observed Yᵢⱼ last value carried forward

1. Known at every event time

In practice the last observation is carried forward between visits, so the value used is out of date.

2. Measured without error

Noisy values plugged in pull the association towards zero (Prentice 1982).

3. External to the event

An internal marker exists only while the patient is event-free, so S(t) cannot be read off the model as a survival probability.

hᵢ(t) = h₀(t) exp{ γᵀwᵢ + α yᵢ(t) },   yᵢ(t): last observed marker value before t

Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, 2nd ed. Wiley 2002. Prentice RL. Biometrika 1982; 69: 331–342.

How much measurement error costs: a simulation

The true log hazard ratio per unit of the marker is 0.6. Entering yearly, error-prone measurements into a time-varying Cox model halves it once the error variance equals the between-patient variance.

00.20.40.60.8 estimated log hazard ratio α̂ true α = 0.6 True current value Yearly, no error Yearly, error SD 0.5 Yearly, error SD 1.0 Yearly, error SD 1.5 0.590.620.520.350.21

Reliability ratio

λ = Var(b₀ᵢ) / {Var(b₀ᵢ) + σ²} is 0.80, 0.50 and 0.31 for error SD 0.5, 1.0 and 1.5; α̂ falls roughly in proportion.

More data does not help

The bias is in the estimand the model targets, not in the sampling error; a larger cohort gives a tighter interval around the wrong value.

mᵢ(t) = b₀ᵢ + (0.3 + b₁ᵢ) t,  b₀ᵢ ~ N(0, 1),  b₁ᵢ ~ N(0, 0.25²);   hᵢ(t) = 0.1 exp{0.6 mᵢ(t)};   Yᵢⱼ = mᵢ(tᵢⱼ) + εᵢⱼ

Illustrative simulation prepared for this talk: n = 1,000 per replicate, 60 replicates per scenario (30 for the reference), yearly visits, censoring at min(8, U(2, 12)) years; dots are means, lines are 2.5th–97.5th percentiles. Reference uses the true mᵢ(t) on a 0.1-year grid.

The joint model: two submodels that share the patient

One submodel follows each patient's marker path; the other describes the event hazard. They share the patient's random effects, so a patient whose marker runs high carries the matching risk.

Longitudinal submodel population path + this patient's deviation + measurement error Survival submodel baseline hazard × baseline covariates × true marker now random effects bᵢ

Yᵢ(t) = mᵢ(t) + εᵢ(t),  mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ

hᵢ(t) = h₀(t) exp{ Wᵢᵀγ + α mᵢ(t) }

Yᵢ(t) observed marker; mᵢ(t) true, error-free marker; bᵢ random effects; εᵢ(t) measurement error; Wᵢ baseline covariates with effects γ; h₀(t) baseline hazard; α association, the log hazard ratio per unit of the true marker.

Why a joint model is not a time-varying Cox model

What the time-varying Cox model sees time value frozen until next visit What the joint model sees time true path mᵢ(t) with uncertainty, at every t

hᵢ(t) = h₀(t) exp{ γᵀwᵢ + α yᵢ(t) },  yᵢ(t) = last observed Yᵢⱼ

hᵢ(t) = h₀(t) exp{ γᵀwᵢ + α mᵢ(t) },  mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ

External versus internal covariate

A covariate is external if its path is unaffected by the event:

Pr{T* ∈ [t, t+dt) | T* ≥ t, 𝒴(t)} = Pr{T* ∈ [t, t+dt) | T* ≥ t, 𝒴(∞)}

A biomarker is internal: it exists only while the patient is event-free, so Pr(T* > t | 𝒴(t)) = 1 and the Cox model gives no survival function.

What the joint model changes

The marker becomes an outcome with its own model, the hazard uses its error-free path, and survival follows by integrating over the patient's random effects bᵢ.

𝒴(t): marker history up to t; 𝒴(∞): the whole path. Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, 2nd ed. Wiley 2002. Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data. CRC Press 2012.

Four ways to use repeated measurements, compared

Time-varying CoxTwo-stageLandmarkingJoint model
How the marker enters the hazardLast observed valueFitted path from a separate mixed modelValue at landmark t; Cox fitted on those at risk at tTrue path mᵢ(t) through shared bᵢ
Between visitsCarried forwardInterpolated by the fitted modelFrozen at tContinuous, with uncertainty
Measurement errorIgnored: α attenuatedPartly removedIgnoredModelled through εᵢ(t)
Dropout caused by the eventIgnoredIgnored in the first stageConditioned on survival to tModelled through bᵢ
Marker uncertainty in the hazardNoNo: standard errors too smallNoYes
Survival probability for an internal markerNot definedDefined, but ignores dropoutDefined from t onwardsDefined at every t

Two-stage: fit the mixed model first, then plug its fitted path into a Cox model. Landmarking: at each prediction time t, a new Cox model among patients still at risk, using the value at t.

Tsiatis AA, DeGruttola V, Wulfsohn MS. JASA 1995; 90: 27–37. van Houwelingen HC. Scand J Stat 2007; 34: 70–85. Tsiatis AA, Davidian M. Stat Sin 2004; 14: 809–834. Rizopoulos D. CRC Press 2012.

Which feature of the marker drives risk?

The link between marker and event is a clinical question: the level today, how fast it is changing, how much has built up, or how far the patient sits from the average.

t

Current value

Risk depends on the true level now.

α mᵢ(t)

Value and slope

Risk also depends on how fast it is changing.

α₁ mᵢ(t) + α₂ mᵢ′(t)

Cumulative effect

Risk depends on accumulated exposure.

α ∫₀ᵗ mᵢ(s) ds

bᵢ

Patient deviation

Risk depends on the patient's departure from average.

αᵀbᵢ

Each term enters the survival submodel as hᵢ(t) = h₀(t) exp{ Wᵢᵀγ + term }, a straight-line effect on the log-hazard scale, an assumption revisited later in the talk.

Estimation: one likelihood, integrated over the patient

Given the random effects, the marker values and the event are independent. The joint likelihood for each patient integrates that product over the random-effects distribution.

p(Yᵢ, Tᵢ, δᵢ; θ) = ∫ { ∏ⱼ p(Yᵢⱼ | bᵢ; θ) } p(Tᵢ, δᵢ | bᵢ; θ) p(bᵢ; θ) dbᵢ

p(Tᵢ, δᵢ | bᵢ) = hᵢ(Tᵢ | ℳᵢ(Tᵢ))^δᵢ · exp{ − ∫₀^Tᵢ hᵢ(s | ℳᵢ(s)) ds }

Conditional independence

Given bᵢ, repeated measurements are independent of each other and of the event. ℳᵢ(t) is the true marker history up to t.

Two integrals

Over bᵢ: Gauss–Hermite or adaptive quadrature, or Laplace approximation. Over time, in the cumulative hazard: Gauss–Kronrod quadrature.

Fitting

EM (Wulfsohn & Tsiatis 1997); MCMC (Faucett & Thomas 1996); Hamiltonian Monte Carlo. Cost grows with patients × quadrature points × random effects.

θ: all parameters (β, D, σ², γ, α, h₀). δᵢ: event indicator. The last card's cost is the scale problem addressed later by jmBIG.

Wulfsohn MS, Tsiatis AA. Biometrics 1997; 53: 330–339. Faucett CL, Thomas DC. Stat Med 1996; 15: 1663–1685. Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data. CRC Press 2012.

Dynamic prediction updates with every visit

At each visit, the patient's own measurements so far update her probability of staying event-free over the window ahead.

follow-up time marker event-free probability new visit t u before the new visit after the new visit

Landmark t, horizon u

The prediction is made at time t, from everything recorded so far, for the window t to u.

Every measurement counts

The whole history to t enters through the posterior of the patient's random effects, with its uncertainty.

Updates at each visit

A new value shifts the random effects and so the whole predicted curve.

πᵢ(u | t) = Pr(Tᵢ* > u | Tᵢ* > t, 𝒴ᵢ(t)) = ∫ {Sᵢ(u | bᵢ) / Sᵢ(t | bᵢ)} p(bᵢ | Tᵢ* > t, 𝒴ᵢ(t)) dbᵢ

𝒴ᵢ(t): marker values recorded up to t; Sᵢ(· | bᵢ): survival function implied by the survival submodel. Rizopoulos D. Biometrics 2011; 67: 819–829.

What a dynamic prediction needs, and who provides it

RequirementBaseline CoxTime-varying CoxLandmarkingJoint model
Updates as new measurements arrive✗✓✓✓
Uses the whole marker history✗✗✗✓
One coherent model across all prediction times✓✓✗✓
Also predicts future marker values✗✗✗✓

Landmarking: a separate Cox model fitted at each prediction time t, among patients still at risk, using the last value. Coherent: predictions at different t come from one set of parameters.

Kalbfleisch JD, Prentice RL. 2nd ed. Wiley 2002. van Houwelingen HC. Scand J Stat 2007; 34: 70–85. Rizopoulos D. Biometrics 2011; 67: 819–829. Putter H, van Houwelingen HC. Landmarking 2.0. arXiv:2103.16252.

Checking a dynamic prediction

Each check is made at a landmark t for a window Δt, among patients still event-free at t, with inverse probability of censoring weights for those censored inside the window.

t t + Δt prediction window patient i: event in window risk 0.42 patient j: event-free past t + Δt risk 0.18

Discrimination

Probability that a patient with the event in the window was given the higher risk.

AUC(t, Δt) = Pr(rᵢ > rⱼ | t < Tᵢ ≤ t+Δt, Tⱼ > t+Δt)

Prediction error

Mean squared difference between predicted and observed status at t + Δt.

BS(t, Δt) = E{ I(Tᵢ > t+Δt) − πᵢ(t+Δt | t) }²

Calibration

Regress observed outcome on the logit of predicted risk: slope 1 and intercept 0 are ideal.

logit Pr(event) = α_cal + β_cal · logit rᵢ

Integrated Brier score

BS averaged over the window, one number per model for comparison.

IBS = (1/Δt) ∫ BS(t, s) ds

rᵢ = 1 − πᵢ(t+Δt | t), the predicted risk in the window; the two risk values in the picture are illustrative. Rizopoulos D. Biometrics 2011; 67: 819–829. Rizopoulos D. J Stat Softw 2016; 72(7).

Dimitris Rizopoulos: joint models into practice

Software anyone can use, predictions that update per patient, ways to judge and combine them, and their use to schedule clinical decisions.

2010 JMjoint models,maximum likelihood 2011 Dynamic predictionspatient-specific survivalupdated with each value 2012 The textbookJoint Models for Longitudinaland Time-to-Event Data 2014 Model averagingBayesian weights that changeby patient and by time 2016 JMbayesBayesian fitting, prediction,discrimination, calibration 2019 Personalised schedulesbiopsy timing in prostatecancer surveillance (PRIAS) Now JMbayes2mixed-type markers,competing risks, multi-state

Rizopoulos D. J Stat Softw 2010; 35(9). Biometrics 2011; 67: 819–829. CRC Press 2012. Rizopoulos D et al. JASA 2014; 109: 1385–1397. Rizopoulos D. J Stat Softw 2016; 72(7). Tomer A et al. Biometrics 2019; 75: 153–162.

Three open issues in the standard joint model

Each one appeared in our patient's chart, and each one weakens the prediction made for her.

1

Scale

Two numerical integrals per patient per iteration. Fits that take minutes on a trial take days on a health-record cohort, so joint models have been confined to small and moderate studies.

Partitioned estimation: jmBIG, 2024

2

Direction

The marker drives the hazard, but a non-terminal event never changes the marker that follows. A post-event jump is absorbed into random-effect variance.

Feedback terms: JMbdirect, 2026

3

Shape

Each unit of the marker adds the same log-risk. Thresholds and plateaus are averaged into one slope, and risks are pulled towards the middle.

Flexible links: JMbdirect, 2026

Computational burden as the barrier: Li S et al. Comput Math Methods Med 2022; 1362913. Papageorgiou G et al. Annu Rev Stat Appl 2019; 6: 223–240.

jmBIG: split by patient, fit in parallel, combine

The cohort is split into patient subsets. Each subset keeps every visit of its patients with their event record, the joint model is fitted to each subset in parallel, and the estimates are combined into one model for prediction.

Full cohort N patients, all visits and events Subset 1 Subset 2 ⋯ Subset S joint model → θ̂₁ joint model → θ̂₂ ⋯ joint model → θ̂ₛ Combined θ̄ Dynamic prediction

θ̄ = (1/S) Σₛ θ̂ₛ,   SE(θ̄) = (1/S) Σₛ SE(θ̂ₛ)

Bayesian routes: posterior means and standard errors averaged over the S subsets. Likelihood route: coefficients and Hessian updated from all subsets together. A patient's rows are never split across subsets.

Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. jmBIG: enhancing dynamic risk prediction and personalized medicine through joint modeling of longitudinal and survival data in big routinely collected data. BMC Med Res Methodol 2024; 24: 172. doi:10.1186/s12874-024-02289-0

One million patients in under an hour

Standard workstation (32 GB RAM) Trusted research environment, health records 01020304050 01020304050 minutes minutes 10100,0001,000,000 10100,0001,000,000 patients (log scale) patients (log scale) 25.6513.0411.215.03 45.2329.2420.5417.49

● Bayesian MCMC (JMbayes2 engine)

● Semiparametric EM (FastJM engine)

● Likelihood EM (joineRML engine)

● Bayesian HMC (Stan engine)

Processing time in minutes by number of patients; values at 1,000,000 patients labelled. A cohort of 1,000 patients with repeated visits and time to death took 2.3, 1.2, 1.2 and 2.72 minutes respectively. Source: Bhattacharjee, Rajbongshi and Vishwakarma, BMC Med Res Methodol 2024; 24: 172, Table 1.

From one direction to both directions

Standard joint model Biomarker path Event hazard α The event never changes the marker path Bidirectional joint model Biomarker path Non-terminal event e.g. cardiovascular Terminal event e.g. death κ₀, κ₁ f₁ f₂ Events feed back into the marker path

Two changes to the standard model: events along the way change the marker that follows (κ₀, κ₁), and the link from marker to each hazard is allowed to bend (f₁, f₂).

Standard one-direction model: Wulfsohn MS, Tsiatis AA. Biometrics 1997; 53: 330–339. Bidirectional model: Bhattacharjee A. JMbdirect: joint modeling with semiparametric link functions for bidirectional feedback in longitudinal-survival data. BMC Med Res Methodol 2026; 26: 230. doi:10.1186/s12874-026-02955-5

Feedback: each event moves the marker's level and slope

A cardiovascular event can shift HbA1c for good; cancer recurrence makes CA-125 jump. The marker submodel gains a level shift and a slope change at each non-terminal event.

time t marker mᵢ(t) Tᵢⱼ κⱼ₀ level shift slope + κⱼ₁ path without the event non-terminal event j

Level shift κⱼ₀

An immediate jump once event j has happened.

Slope change κⱼ₁

Faster or slower progression after the event.

What the data must show

At least one measurement after the event for the jump and two for the slope; about five such patients per event type across the cohort.

mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ + Σj ∈ 𝒥ᵢ(t) { κⱼ₀ I(t ≥ Tᵢⱼ) + κⱼ₁ (t − Tᵢⱼ)₊ }

𝒥ᵢ(t): non-terminal events patient i has had by t; Tᵢⱼ: time of event j; (x)₊ = max(x, 0). Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

Shape: risk does not always rise in a straight line

The standard model adds the same log-risk for each unit of the marker. Real risk often has a threshold or a plateau. A smooth function in place of the straight line lets the data show the shape.

true marker value m log-risk f(m) threshold straight line α m flexible f(m)

Penalised spline

Built from B-spline pieces Bₗ joined smoothly; a penalty stops it wiggling more than the data support.

Neural additive component

A sum of Q S-shaped steps σ(·) that can capture sharp thresholds, also penalised.

Still a likelihood

Fitted by penalised likelihood with Laplace approximation, or Bayesian Hamiltonian Monte Carlo, with pointwise uncertainty bands for f.

hᵢₖ(t) = h₀ₖ(t) exp{ Wᵢᵀγₖ + fₖ(mᵢ(t)) }

fₖ(m) = Σₗ αₗ Bₗ(m)  or  Σq wq σ(aq m + bq)

k indexes the event type (competing or recurrent events), each with its own link fₖ. Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

Simulation: feedback and a flexible link predict best

Six specifications fitted to the same simulated cohorts with competing risks. The spline link with feedback ranked patients best and had the smallest prediction error.

AUC at 5 years (higher is better) Integrated Brier score, 0–5 years (lower is better) Linear, no feedbackLinear, feedbackSpline, no feedbackSpline, feedbackNeural, no feedbackNeural, feedback 0.700.750.800.85 0.130.140.150.160.17 0.720.750.820.850.810.84 0.1680.1560.1410.1340.1430.136

● no feedback term   ● feedback term included. Feedback adds to every link; the flexible link adds more.

Coverage of 95% intervals: 0.88 for linear without feedback, 0.95 for spline with feedback.

n = 500, competing events, piecewise-constant baseline hazard; means over 1,000 replicates. AUC and integrated Brier score as defined on the prediction-checking slide. Source: Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230, Table 6.

Calibration, and the comparison with established joint models

The straight-line model without feedback compresses risks towards the middle (slope 0.74). With feedback and a spline link, predicted risks match what happens (slope 0.99).

Calibration slope β_cal (ideal 1) and intercept α_cal AUC at 5 years on identical data 0.700.800.901.00 0.600.700.800.90 Linear, no feedbackLinear, feedbackSpline, no feedbackSpline, feedbackNeural, no feedbackNeural, feedback Two-stagejoineRMLJMbayes2JMbdirect 0.740.820.930.990.920.98 α 0.18α 0.12α 0.06α 0.02α 0.07α 0.03 0.670.710.730.85 Comparators: linear link, no feedback.

n = 500, competing risks, non-linear true association, moderate feedback, landmark t = 0, horizon 5 years; means over 1,000 replicates with the same training and test splits. Source: Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230, Tables 5 and 9. JMbdirect here: spline link with feedback.

One workflow from model to prediction

Biomarker and event histories JMbdirect feedback κ and flexible links fₖ jmBIG pathway partitioned, parallel estimation Individual dynamic πᵢₖ(u | t)

Why the routes are separate

Feedback terms and flexible links need each patient's whole trajectory and use their own routines: splines in the M-step, neural-component optimisation and feedback-augmented trajectory updates.

One specification

The same model specification routes to the partitioned pathway when the cohort is large, and predictions return in the same form, πᵢₖ(u | t) for event type k.

Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. BMC Med Res Methodol 2024; 24: 172. Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

Where joint models are heading

Many markers, many diseases

Multimorbidity as a network: each hazard driven by several trajectories, each event reshaping several markers. The multivariate extension of the bidirectional model is under development.

Feedback by design

Events that change the biomarker path written into the model, rather than absorbed into random-effect variance.

Health-record scale

Partitioned and parallel estimation inside trusted research environments, so a million-patient cohort fits in under an hour.

Flexible links

Splines and neural additive terms that reveal thresholds and plateaus while keeping a likelihood and uncertainty bands.

Deep-learning survival

Dynamic-DeepHit and recurrent-network models scale to very large cohorts; joint models keep explicit, interpretable structure and pointwise uncertainty.

Prediction in the clinic

Risk that updates at every visit, built into dashboards and health-record systems for patient-level decisions such as surveillance intervals.

Lee C, Yoon J, van der Schaar M. Dynamic-DeepHit. IEEE Trans Biomed Eng 2020; 67: 122–133. Tomer A et al. Biometrics 2019; 75: 153–162.

Take-home messages

1

A repeatedly measured biomarker should not go straight into a Cox model: measurement error alone can halve the estimated association.

2

Joint models fit marker and event together through the patient's random effects and give predictions that update at every visit.

3

jmBIG fits joint models to health-record cohorts of a million patients in under an hour (BMC Med Res Methodol 2024; 24: 172).

4

JMbdirect lets events change the marker and risk follow non-linear shapes, giving sharper, better-calibrated predictions (BMC Med Res Methodol 2026; 26: 230).

References (1)

Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. JASA 1958; 53: 457–481.

Cox DR. Regression models and life-tables. J R Stat Soc B 1972; 34: 187–220.

Rubin DB. Inference and missing data. Biometrika 1976; 63: 581–592.

Laird NM, Ware JH. Random-effects models for longitudinal data. Biometrics 1982; 38: 963–974.

Prentice RL. Covariate measurement errors and parameter estimation in a failure time regression model. Biometrika 1982; 69: 331–342.

Diggle P, Kenward MG. Informative drop-out in longitudinal data analysis. Appl Stat 1994; 43: 49–93.

Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, 2nd ed. Wiley; 2002.

DeGruttola V, Tu XM. Modelling progression of CD4-lymphocyte count and its relationship to survival time. Biometrics 1994; 50: 1003–1014.

Faucett CL, Thomas DC. Simultaneously modelling censored survival data and repeatedly measured covariates: a Gibbs sampling approach. Stat Med 1996; 15: 1663–1685.

Wulfsohn MS, Tsiatis AA. A joint model for survival and longitudinal data measured with error. Biometrics 1997; 53: 330–339.

Henderson R, Diggle P, Dobson A. Joint modelling of longitudinal measurements and event time data. Biostatistics 2000; 1: 465–480.

Tsiatis AA, Davidian M. Joint modeling of longitudinal and time-to-event data: an overview. Stat Sin 2004; 14: 809–834.

Brown ER, Ibrahim JG, DeGruttola V. A flexible B-spline model for multiple longitudinal biomarkers and survival. Biometrics 2005; 61: 64–73.

Chi YY, Ibrahim JG. Joint models for multivariate longitudinal and multivariate survival data. Biometrics 2006; 62: 432–445.

Elashoff RM, Li G, Li N. Joint analysis of longitudinal measurements and competing risks failure time data. Stat Med 2007; 26: 2813–2835.

van Houwelingen HC. Dynamic prediction by landmarking in event history analysis. Scand J Stat 2007; 34: 70–85.

References (2)

Rizopoulos D. JM: an R package for the joint modelling of longitudinal and time-to-event data. J Stat Softw 2010; 35(9): 1–33.

Rizopoulos D. Dynamic predictions and prospective accuracy in joint models. Biometrics 2011; 67: 819–829.

Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data. CRC Press; 2012.

Rizopoulos D, Hatfield LA, Carlin BP, Takkenberg JJM. Combining dynamic predictions using Bayesian model averaging. JASA 2014; 109: 1385–1397.

Proust-Lima C et al. Joint latent class models for longitudinal and time-to-event data: a review. Stat Methods Med Res 2014; 23: 74–90.

Rizopoulos D. The R package JMbayes. J Stat Softw 2016; 72(7): 1–46.

Hickey GL et al. joineRML: a joint model and software package. BMC Med Res Methodol 2018; 18: 50.

Papageorgiou G et al. An overview of joint modeling of time-to-event and longitudinal outcomes. Annu Rev Stat Appl 2019; 6: 223–240.

Tomer A et al. Personalized schedules for surveillance of low-risk prostate cancer patients. Biometrics 2019; 75: 153–162.

Lee C, Yoon J, van der Schaar M. Dynamic-DeepHit. IEEE Trans Biomed Eng 2020; 67: 122–133.

Putter H, van Houwelingen HC. Landmarking 2.0. arXiv:2103.16252.

Li S et al. Efficient algorithms for a semiparametric joint model: massive biobank data. Comput Math Methods Med 2022; 1362913.

Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. jmBIG: enhancing dynamic risk prediction and personalized medicine through joint modeling of longitudinal and survival data in big routinely collected data. BMC Med Res Methodol 2024; 24: 172.

Bhattacharjee A. JMbdirect: joint modeling with semiparametric link functions for bidirectional feedback in longitudinal-survival data. BMC Med Res Methodol 2026; 26: 230.

References (3): history and the last fifteen years

Wu MC, Carroll RJ. Estimation and comparison of changes in the presence of informative right censoring by modeling the censoring process. Biometrics 1988; 44: 175–188.

Wu MC, Bailey KR. Estimation and comparison of changes in the presence of informative right censoring: conditional linear model. Biometrics 1989; 45: 939–955.

DeGruttola V, Lange N, Dafni U. Modeling the progression of HIV infection. JASA 1991; 86: 569–577.

Self S, Pawitan Y. Modeling a marker of disease progression and onset of disease. In: Jewell NP, Dietz K, Farewell VT (eds). AIDS Epidemiology: Methodological Issues. Birkhäuser; 1992.

Pawitan Y, Self S. Modeling disease marker processes in AIDS. JASA 1993; 88: 719–726.

Follmann D, Wu M. An approximate generalized linear model with random effects for informative missing data. Biometrics 1995; 51: 151–168.

Tsiatis AA, DeGruttola V, Wulfsohn MS. Modeling the relationship of survival to longitudinal data measured with error: applications to survival and CD4 counts in patients with AIDS. JASA 1995; 90: 27–37.

Tsiatis AA, Davidian M. A semiparametric estimator for the proportional hazards model with longitudinal covariates measured with error. Biometrika 2001; 88: 447–458.

Song X, Davidian M, Tsiatis AA. A semiparametric likelihood approach to joint modeling of longitudinal and time-to-event data. Biometrics 2002; 58: 742–753.

Rizopoulos D, Verbeke G, Lesaffre E. Fully exponential Laplace approximations for the joint modelling of survival and longitudinal data. J R Stat Soc B 2009; 71: 637–654.

Proust-Lima C, Taylor JMG. Development and validation of a dynamic prognostic tool for prostate cancer recurrence using repeated measures of posttreatment PSA: a joint modeling approach. Biostatistics 2009; 10: 535–549.

Taylor JMG, Park Y, Ankerst DP et al. Real-time individual predictions of prostate cancer recurrence using joint models. Biometrics 2013; 69: 206–213.

Crowther MJ, Abrams KR, Lambert PC. Joint modeling of longitudinal and survival data. Stata J 2013; 13: 165–184.

Gould AL, Boye ME, Crowther MJ et al. Joint modeling of survival and longitudinal non-survival data: current methods and issues. Report of the DIA Bayesian joint modeling working group. Stat Med 2015; 34: 2181–2195.

Rizopoulos D, Molenberghs G, Lesaffre E. Dynamic predictions with time-dependent covariates in survival analysis using joint modeling and landmarking. Biom J 2017; 59: 1261–1276.

Brilleman S, Crowther M, Moreno-Betancur M, Buros Novik J, Wolfe R. Joint longitudinal and time-to-event models via Stan. StanCon 2018.

Lin J, Luo S. Deep learning for the dynamic prediction of multivariate longitudinal and survival data. Stat Med 2022; 41: 2894–2907.

Rustand D, van Niekerk J, Krainski ET, Rue H, Proust-Lima C. Fast and flexible inference for joint models of multivariate longitudinal and survival data using INLA. arXiv:2203.06256.