Patients often heal on the scan without feeling better, or feel better while their tissue keeps failing. A knee repair can relieve pain while balance stays impaired. A hip patient can feel ready for discharge without being any more able to care for themselves. Most clinical tools miss this mismatch for three reasons. Outcomes are measured at one visit. Objective and patient-reported measures are modelled separately. The bedside risk calculators clinicians actually use, called nomograms (paper charts where each predictor maps to a number of points, and the summed points map to a probability), take only frozen snapshot values and return one probability.
We propose to learn patient subgroups from the joint trajectory of an objective tissue marker and a patient-reported symptom score. Each subgroup is defined by how the two curves move together over time, including which one moves first. We then compress early-visit membership in each subgroup into a nomogram whose axes are levels and rates of change of both measures. The claim to test is specific. Coupled trajectory subgroups carry prognostic information that simpler options do not capture. Those simpler options are two-visit change-score quadrants (structure better or worse, crossed with symptoms better or worse) and supervised slope-based nomograms. If the claim fails, the result is still useful, because it tells clinicians that the cheap quadrant is enough.
Core Claim
The closest prior model is the Probabilistic Subtyping Model of Clustering Longitudinal Clinical Marker Trajectories from Electronic Health Data. It clusters irregularly sampled marker trajectories using four layers: subtype-level B-spline mean curves, covariate regressions, individual random intercepts and slopes, and a Gaussian-process term that absorbs short-term episodic noise. Applied to scleroderma, it found a recovering lung-function endotype and cut forecast error for percent-predicted forced vital capacity from 9.27 to 7.69 RMSE.
That model is univariate: each marker is clustered on its own. We supply its method but change the object being clustered. Each subtype becomes a pair of mean curves, one for structure and one for symptoms, and the individual deviations are correlated across the two. A subtype is therefore defined by the shape of the structure–symptom relationship, not by either curve alone.
Concretely, for patient $i$ in subtype $k$ at time $t$, structure $y^{s}_{i}(t)$ and symptom score $y^{p}_{i}(t)$ are modelled as
$$ y^{m}_{i}(t) = \phi(t)^\top \beta^{m}_{k} + x_i^\top \gamma^{m} + b^{m}_{i0} + b^{m}_{i1} t + f^{m}_{i}(t) + \epsilon^{m}_{it}, \quad m \in \{s, p\}. $$
The terms are as follows. $\phi(t)$ is a B-spline basis, so $\beta^{m}_{k}$ gives subtype $k$'s mean curve for modality $m$. $x_i$ are baseline covariates with effects $\gamma^m$. The four random intercepts and slopes $(b^{s}_{i0}, b^{s}_{i1}, b^{p}_{i0}, b^{p}_{i1})$ share one full 4×4 covariance per subtype, $\Sigma_k$, which is the cross-modal coupling. $f^{m}_{i}$ is a per-modality Gaussian process for short-term fluctuations, and $\epsilon$ is measurement noise.
The coupling in $\Sigma_k$, together with offsets between the two subtype mean curves, is what lets the model represent a lead–lag pattern such as "cartilage changes first, pain follows a year later." A quadrant of change scores cannot represent that pattern.
Research Program
The primary setting is the Osteoarthritis Initiative (OAI): 48 months of 3D knee MRI in a multicenter cohort. Charting Aging Trajectories of Knee Cartilage Thickness segmented cartilage with a 3D UNet, built age–BMI centile curves with generalized additive mixed models, and showed that early disease involves transient cartilage thickening before loss. That finding motivates this proposal. If structure can move up before it moves down, the timing of symptoms relative to that swing is exactly what a snapshot misses. The same paper also constrains the proposal: its supervised LightGBM on centile scores is the strongest existing predictor and has to be beaten or matched. We reuse its centile scores as the structural input.
For the symptom input, the research file lists only imaging, Kellgren–Lawrence grade (KLG, the standard 0–4 radiographic osteoarthritis grade) and demographics for OAI. The proposal therefore depends on confirming that the release we obtain contains a repeated knee-symptom questionnaire. If it does not, the project relocates to a cohort that does, and the scleroderma and stroke data below are used only for the method checks.
The model is fit by expectation–maximization, extending the PSM inference to the bivariate case. The number of subtypes is chosen by held-out likelihood, not by eye.
The interpretation lens comes from multi-dimensional outcome work showing that objective and subjective axes can come apart:
- the proprioceptive gap after mosaicplasty, reported in Effect of autogenous osteochondral mosaicplasty on the balance control of patients with cartilage defects of the knee;
- motor deficits predicting post-stroke disability better than pain intensity, reported in Multifaceted evaluation of central post–stroke pain;
- discharge readiness tracking Harris Hip Score (4.002 points higher, p = 0.004) but not daily-living ability (p = 0.888), reported in Study on the effect factors of discharge readiness of total hip arthroplasty patients.
Those studies are cross-sectional or have a single follow-up. We ask whether the mismatch has a reproducible temporal shape.
The distillation step turns subtypes into a bedside chart. For each clinically important subtype, for example "structural progression with silent symptoms," we fit a sparse logistic model. Its predictors are the patient's first two or three visits summarized as levels and per-year slopes of both modalities, plus their product. That model is drawn as a standard nomogram, with each predictor's points scaled relative to the largest coefficient, as in A nomogram to predict the risk of prolonged length of stay following primary total hip arthroplasty. That work used only frozen day-1 values. Its "dynamic" relative in A novel visual dynamic nomogram to online predict the risk of unfavorable outcome in elderly aSAH patients is a web calculator over admission snapshots. Our chart differs in that its axes are rates of change, and in that its target is a data-driven subtype rather than a dichotomized endpoint.
Fidelity is measured by how closely the chart reproduces the full model's subtype probabilities on held-out patients.
Two secondary cohorts check the method:
- On the scleroderma registry, we first check that our univariate special case reproduces the published PSM error. We then fit pairs of markers to see whether coupling improves forecasting.
- The NINDS tPA stroke trial gives five repeated visits of NIH Stroke Scale subscores. Stroke recovery phenotyping through network trajectory approaches and graph neural networks found three recovery subtypes there with network clustering, and found that a graph neural network predicts subtype membership early. We test whether a two-axis motor-versus-sensory coupled model recovers comparable subtypes, and whether a slope nomogram from the 2- and 24-hour visits matches the graph network's early prediction. That paper reports no numeric accuracy, so it is a qualitative target only.
Baselines
The naive reference is a static baseline nomogram. It uses age, sex, BMI, baseline KLG, baseline cartilage centile and baseline symptom score in a logistic model predicting the 48-month outcomes. These outcomes are radiographic incidence and the joint state "structural progression with or without symptomatic worsening." This baseline represents current nomogram practice, and every other model must beat it to justify needing repeat visits.
The strongest existing method is the centile-score LightGBM from Charting Aging Trajectories of Knee Cartilage Thickness. It reports AUC = 0.81 (F1 = 0.69) for incidence of KLG < 2 progressing to ≥ 2, and AUC = 0.64 for the KLG 0 to 1 transition. It is the strongest competitor because it already uses longitudinal structural information with a flexible learner. We rerun it on identical splits, with symptom scores added as features so the comparison is not handicapped. We give it a hyperparameter search budget equal to the one we spend on choosing the number of subtypes and the spline knots.
"Beating it" is defined by intended use. The distilled nomogram must come within 0.02 AUC of LightGBM on incidence while being hand-computable. The full coupled model must exceed LightGBM on the joint structure–symptom outcome, for which no reported result was found on Lacuna. Literature numbers for related tasks are context only. Examples are the hip length-of-stay nomogram's C-index of 0.863 and the 0.882 AUC of the subarachnoid haemorrhage nomogram; neither is a target on OAI.
Two further baselines isolate the key ingredient, cross-modal coupling:
- Separate univariate PSMs, one fit to structure and one to symptoms, with patients cross-classified by their two assignments. This tests whether joint clustering finds anything that independent clustering does not.
- Coupling ablation. Our model with $\Sigma_k$ restricted to block-diagonal, so the subtype mean curves are shared but individual deviations are independent across modalities.
A third ablation removes the Gaussian-process noise layer. This tests whether short-term symptom flares create spurious subtypes, which is the failure mode PSM was designed to prevent.
Why not just…?
Why not just split patients into four quadrants by two-visit change in structure and symptoms? This is the most serious threat, and the proposal is built to settle it. The quadrant is cheap and transparent, and it already encodes discordance. It cannot express a lagged or non-monotone relationship such as the thickening-then-loss pattern on the structural side. We will compare held-out 48-month prediction and held-out trajectory likelihood between the quadrant model (given the same visits) and the coupled model. If the quadrant matches within noise, we report that and recommend the quadrant chart. In that case the paper's contribution becomes a validated two-axis change nomogram rather than new subtypes.
Why not just fit a supervised nomogram on slopes, with no clustering? A slope-feature logistic model predicting each 48-month outcome directly answers "what is this patient's risk" without subtypes. Two things can separate it from our approach. First, whether subtype probabilities improve calibration for the joint outcome. Second, whether the subtypes hold up across OAI's clinical sites. Unsupervised clusters are known to shift between sites, a limitation the low back pain clustering of An Unsupervised Learning Approach for Multimodal Low Back Pain Stratification also inherits from being cross-sectional. We include this model as a baseline. If it matches the subtype-based charts on both checks, the subtypes have no prognostic value, although they may still have descriptive value.
Why not just add symptom scores to LightGBM and stop? Digital health technologies and machine learning augment patient reported outcomes shows that combining objective and patient-reported streams helps: in rheumatoid arthritis, sensor features plus questionnaires gave F1 = 0.833 for severity versus 0.759 for questionnaires alone. A boosted model with both inputs may simply win on discrimination. It cannot be computed by hand, however, and it does not say which pattern of mismatch a patient follows. We measure the gap between it and the distilled nomogram directly rather than assume the gap is small.
Proposed Experiments
The decisive check is a held-out comparison on OAI patients split by site. It compares the coupled model, the coupling ablation, separate univariate PSMs, quadrants, and the slope nomogram. Metrics are trajectory forecast error for both modalities at 36–48 months given visits up to month 24, and AUC and calibration for the 48-month outcomes. The coupled model has to win on the joint outcome and on symptom forecasting to support the claim.
Subtype stability is measured with leave-one-site-out refits. We match subtypes across refits and report the adjusted Rand index. Subtypes with low agreement are not drawn as nomograms.
Nomogram fidelity is reported three ways: the C-index of the chart against the full model's assignments, the AUC loss relative to full-model posteriors, and how often hand-computed probabilities differ from software output, tested in a small clinician exercise.
On the scleroderma registry, we must reproduce RMSE near 7.69 for percent-predicted forced vital capacity before trusting the bivariate extension.
Finally, we check substantively whether a subtype appears in which structural thickening precedes symptom change. Such a subtype is the pattern the centile study implies, and the kind of finding that would justify modelling the two curves jointly.
References
- Clustering Longitudinal Clinical Marker Trajectories from Electronic Health Data: Applications to Phenotyping and Endotype Discovery
- Charting Aging Trajectories of Knee Cartilage Thickness for Early Osteoarthritis Risk Prediction: An MRI Study from the Osteoarthritis Initiative Cohort.
- Effect of autogenous osteochondral mosaicplasty on the balance control of patients with cartilage defects of the knee: a pilot study
- Multifaceted evaluation of central post–stroke pain
- Study on the effect factors of discharge readiness of total hip arthroplasty patients
- A nomogram to predict the risk of prolonged length of stay following primary total hip arthroplasty with an enhanced recovery after surgery program
- A novel visual dynamic nomogram to online predict the risk of unfavorable outcome in elderly aSAH patients after endovascular coiling: A retrospective study
- Stroke recovery phenotyping through network trajectory approaches and graph neural networks
- An Unsupervised Learning Approach for Multimodal Low Back Pain Stratification
- Digital health technologies and machine learning augment patient reported outcomes to remotely characterise rheumatoid arthritis