Can latent similarity Gaussian processes personalize suicide-risk forecasting?

Pooled and per-patient models each assume too much

The paper frames suicide-risk forecasting as a choice between three existing model classes, each carrying an assumption that limits clinical use. Nomothetic (pooled) models share one global trend across all patients, which treats any unobserved, patient-specific variation as uncorrelated noise; that is only safe if every meaningful driver of risk has been observed. Idiographic (per-patient) models fit each patient independently, discarding shared population dynamics and, because suicide attempts occur at a low base rate, are prone to overfitting when a patient's data is sparse. Hierarchical (mixed-effects) models decompose forecasts into a shared population trend (fixed effects) and patient-specific deviations (random effects), but assume those deviations are random draws from a single global trend.

The paper argues this last assumption is clinically critical. Patients are described as inhabiting a continuous spectrum of dynamic subtypes governed by distinct, potentially conflicting trends. Under sparse event histories, hierarchical models force risk estimates for patients with no prior attempts to default to the global population trend, which may represent them poorly. The paper notes that a substantial proportion of suicide fatalities occur on the first attempt, so a model that defaults first-attempt patients to the population average is failing exactly where it matters.

A similarity kernel lets patients borrow from latent peers

LSGPs place patients in a continuous latent space and jointly model similarity and risk. When a patient's own data is limited, the model draws information from similar patients to forecast that patient's future risk. The similarity factor is defined as a function of a weighted feature distance between latent codes, and the paper shows that many popular kernels (RBF, Matérn 5/2, rational-quadratic) are instantiations of this formulation, each combinable with any code-independent input kernel.

The paper extends these stationary, constant-amplitude kernels with two mechanisms: an input-dependent amplitude, where the amplitude is a softplus of a neural network evaluated at the latent code, and non-stationarity via per-dimension input warping. The warping uses a softplus parameterization that guarantees positivity by construction, which the paper states makes the transformation strictly order-preserving and satisfies the kernel definition. The paper also proves the two-channel similarity kernel is identifiable: the trend and amplitude components are pinned down up to a positive scale factor.

LSGPs relax four constraints of hierarchical models

The paper recalls that hierarchical (mixed-effects) models can be written as Gaussian processes with a linear fixed-effect kernel plus an indicator-gated linear deviation kernel. The indicator term I(n = m) is what makes a patient's random effect apply only to that patient.

LSGPs generalize this in four specific ways: they replace the rigid indicator I(n = m) with a learned continuous similarity kernel between latent codes; they replace the fixed identity input n with an inferable low-dimensional code; they relax linear feature kernels to flexible input kernels; and they parameterize shrinkage strength via a learned, patient-specific amplitude function produced by a neural network. This is the paper's own account of the relationship, not an empirical claim, and it is what grounds the claim that LSGPs generalize nomothetic, idiographic, and hierarchical frameworks.

Mean-field inference silently collapses LSGPs to pooled models

The paper's second contribution is a negative result about the standard fitting algorithm. Mean-field variational inference factorizes the posterior as q(Z, U) = q(Z) · q(U), treating the inducing values U as independent of the latent codes Z. The paper proves that this factorization adds a penalty to the objective, which it calls the coupling gap, equal to the Bregman information of the natural parameters of the conditional posterior.

The gap is zero if and only if those natural parameters are almost surely constant in the codes. The paper then proves that a zero coupling gap forces what it calls personalization collapse: the similarity kernel's importance weights go to zero, the amplitude becomes constant, and the combined similarity matrix becomes identically one. In other words, the objective that mean-field inference actually optimizes is the structured objective minus a term that rewards exactly the code-dependence the model needs. The paper notes this echoes a known asymptotic bias in Bayesian neural networks with latent variables, where mean-field inference underfits and miscalibrates uncertainty.

A Pólya-Gamma scheme keeps the code-dependent posterior

The fix adapts a Pólya-Gamma augmentation scheme to the LSGP. The Bernoulli likelihood is rewritten as an expectation over an auxiliary variable, which restores conditional Gaussian conjugacy for the inducing values U. This yields a structured variational family in which the inducing posterior q(U | Z) depends on the latent codes, rather than being factorized away from them.

The paper then derives closed-form coordinate-ascent updates for the conjugate factors: an auxiliary update for the local Pólya-Gamma factors, and an inducing-posterior update for the global factor conditioned on Z. Parameters are updated in a separate stochastic gradient step. For each outer variational expectation-maximization step, the two conjugate updates are repeated for 15 sweeps to reach conditional convergence; with mini-batching and unit step size, those conjugate updates are exactly natural-gradient steps. Only the code posterior q(Z; ϕ) needs to be parameterized, and it is amortized through a shared encoder built as a DeepSet over each patient's unordered observation set, which guarantees permutation invariance by construction.

Gains are largest for first-occurrence events

The empirical results come from intensive longitudinal suicide data. LSGPs outperform nomothetic, idiographic, and hierarchical models for next-week risk forecasting, and the largest gains appear when forecasting first-occurrence suicide-related events (SREs). The paper's limitations section restates this in patient terms: within-patient forecasts improve over all baselines, with the largest gains in patients who have no training-window SREs and who go on to experience an SRE in the test window.

That subgroup is the one the introduction flagged as the failure mode of hierarchical models, since those patients have no prior attempts and would otherwise default to the global population trend. The paper also reports that static demographic traits do not account for the learned similarity structure, and it interprets the latent geometry as likely capturing underlying, unobserved clinical dynamics. That interpretation is the authors' reading of the latent space, not a measured correspondence to clinical phenotypes; the paper explicitly lists validating the learned similarities against richer clinical variables as future work.

The log-loss metric itself is called into question

The paper reports that log loss fails to differentiate uninformative base-rate models from predictive models on the EMA data. Because of this, the authors question its suitability for model selection and its impact on Bayesian inference, and they state a plan to investigate alternative training objectives and selection criteria that optimize PAC-Bayes generalization bounds rather than relying on marginal likelihoods.

This matters for reading the headline comparison. The claim that LSGPs outperform the baselines is made on this data, and the paper itself flags that one of the metrics used in this setting does not separate a base-rate model from a predictive one. The paper also notes that the current LSGP evaluates risk cross-sectionally via latent codes, omitting temporal dynamics, which is a stated open problem rather than a solved feature.

The evidence is one intensive longitudinal cohort

The empirical support comes from a single intensive longitudinal suicide dataset, with next-week risk as the forecast horizon and SREs as the outcome. The paper's own framing of the three model classes, the identifiability proof, and the personalization-collapse theorem are theoretical results about the model class and the inference algorithm, not empirical findings about clinical populations. The performance comparison is the empirical part, and it is bounded to this data, this horizon, and this outcome.

The clinical value the paper claims for the similarity structure is that it exposes relevant historical precedents that may prompt clinicians to investigate unmeasured contextual factors driving risk. That is a proposed use of the model's output, and the paper treats validating the learned similarities against richer clinical variables as future work. The paper also used AI for initial proof sketching, then rewrote, finished, and verified all proofs, and computed every reported numerical result directly from raw experimental artifacts.

Sources used in this answer