Dynamic learning of individual-level suicidal ideation trajectories to enhance mental health care

Routine outcome monitoring using digital tools underpins measurement-based care (MBC), aiming to enhance clinical decision-making and personalize treatments. Despite its benefits, clinicians’ predictive performance is often poor, highlighting the need for models that can predict future states and update dynamically as new data arrive. Many existing predictive approaches produce categorical predictions over fixed horizons and commonly ignore rich longitudinal histories or irregular sampling. In suicidality, categorical prediction performance (e.g., positive predictive value) remains low despite known associations with prior suicidality, diagnoses, symptoms, functioning, and demographics. Suicidal ideation often fluctuates over short intervals, necessitating trajectory-focused approaches that can accommodate irregular observation times, individual heterogeneity, bounded scales, and uncertainty quantification. The authors propose a continuous-time modeling framework, well-suited to the naturalistic, irregular data of MBC, to learn and update individual trajectories of suicidal ideation and support ongoing clinical decision-making.

Prior work shows MBC improves outcomes, retention, monitoring, and patient–clinician communication, yet translating measurements into individualized predictions remains challenging. Traditional predictive models in mental health typically forecast over fixed periods and rarely incorporate full longitudinal histories or update dynamically. In suicidality research, models achieve only modest gains over chance in positive predictive value, even with known correlates (prior suicidal thoughts/behaviors, diagnoses, symptom profiles, functioning, sociodemographics). High-frequency monitoring studies reveal substantial short-term fluctuations in ideation influenced by mood, hopelessness, loneliness, affective instability, and sleep, underscoring the need to model individual trajectories and variability. Continuous-time models have desirable properties for irregularly spaced clinical data (handling missingness/irregularity, hierarchical structure, Bayesian uncertainty, link functions mapping latent states to bounded scales) but are underused in mental health compared to discrete-time approaches. This work addresses this gap by applying continuous-time stochastic processes to suicidal ideation within an MBC context.

Design and ethics: Approved by Northern Sydney Local Health District Human Research Ethics Committees (HREC/17/HAWKE/480). Participants provided online informed consent (opt-out). Participants and setting: Individuals presenting for the first time to youth mental health services including Headspace sites in NSW, QLD, SA (N=400, 68.4%), Mind Plasticity (N=159, 27.1%), Open Arms (N=13, 2.2%), Butterfly Foundation (N=13, 2.2%), between Nov 2018 and Nov 2022. Inclusion: use of the Innowell platform and at least two completed suicidal ideation assessments. Platform and measures: Innowell is a web-based platform enabling assessment, monitoring, and feedback. Initial assessment covers biopsychosocial domains; follow-ups include SIDAS (suicidal ideation), C-SSRS (behaviors), SOFAS, CGI-S, EQ-5D-Y (overall health), and Schuster’s SSS (social support). The primary outcome is the summed SIDAS score (five items on 0–10 scales; total range 0–50; higher is worse). Baseline SIDAS from the initial assessment; follow-ups collected at variable times. Modeling approach: Continuous-time stochastic differential equation models to handle irregular observation intervals. Two processes were evaluated: (1) Wiener process (continuous-time random walk) with individual diffusion σ_m^2 and baseline φ_m; (2) Ornstein–Uhlenbeck (OU) process adding drift toward a long-term constant θ_m at rate ω_m. Latent state values are mapped to bounded, discretized SIDAS scale using a logistic link and a discretized normal observation model. Parameter estimation uses hierarchical (mixed-effects) structure with population-level priors and individual-level random effects where specified. Inference is Bayesian via posterior sampling: correlated pseudo-marginal Metropolis–Hastings for individual parameters within fixed population parameters, and semi-conjugate Gibbs sampling for population parameters. A particle filter is used for likelihood approximation when closed-form solutions are unavailable. Prediction and summaries: Posterior predictive simulations generate 60-day-ahead trajectories from a reference timepoint, summarized daily by medians and equal-tailed credible intervals (ETIs). Summary statistics include: (a) Integrated High-Ideation Probability (IHIP), probability of being in high ideation (SIDAS > 20) at least once in the next 60 days; (b) Future observational variability (V), the expected ETI range at day 60 from a reference time with 80% highest density intervals; (c) Recommended follow-up time (FUT), number of days until the 95% ETI spans at least half the SIDAS range (i.e., >25), as a proxy for information decay. Model comparison: Predictive performance assessed using WAIC on individuals with ≥10 observations (N=55). Multiple parameterizations with fixed vs random effects were compared to identify the best-performing, parsimonious model. Software: Model code/sampling in Julia 1.8; analyses in R 4.3.1. Code repository: github.sydney.edu.au/mvar0005/dynamiclearning/.

Sample characteristics (N=585): 72.6% female; mean age 24.2 years (SD 10.8). Lifetime suicidal plans/attempts: 40.7%; in last 3 months: 10.4%. Mean number of SIDAS observations per person 4.5 (median 2); 55 individuals had ≥10 observations; 140 had ≥5. Median inter-observation interval 35 days (IQR 1–108). A substantial subset (26.0%) had zero ideation throughout care. Model performance: The Wiener process with random effects for baseline and diffusion (i.e., individual-specific parameters) achieved the best predictive performance (WAIC 5303.8) compared to alternatives including a mixed-effects Wiener model (WAIC 5460.2) and an OU model with additional drift/constant terms (WAIC 5390.4). Allowing individual-specific diffusion markedly improved fit, indicating heterogeneity in variability across individuals. Parameter insights: Two key individual parameters are the transformed baseline (φ_m) and diffusion (σ_m^2). No strong dependency was observed between baseline level and diffusion; many individuals exhibited baseline SIDAS at zero. Diffusion captures volatility; higher values correspond to greater variability in trajectories on the SIDAS scale. Correlates of variability: The diffusion parameter was positively associated with recent (last 3 months) suicidal plans/attempts (mean correlation 0.10; 95% ETI 0.02, 0.19) and showed a borderline association with mania-like experiences (ASRM; mean 0.07; 95% ETI 0.00, 0.14). Other demographic and baseline clinical variables showed weak or null associations. Predictive summaries and monitoring: Across individuals with ≥10 observations, future variability V had a median of 10.2 (IQR 0.0–36.9). FUT had a median of 6 days (IQR 3–19), indicating how quickly information from prior observations degrades and informing suggested monitoring intervals. Individual examples demonstrated dynamic updating: initial uncertainty at baseline narrowed as more observations accrued, with corresponding increases in FUT (e.g., from 2 to 35 days) and decreases in V for individuals with relatively stable courses. Clinical implications: Trajectories differ qualitatively across individuals (e.g., improving, deteriorating, high-risk but stable, high variability, persistent zero ideation). Individualized predictions provide nuanced estimates of near-term high-ideation risk (IHIP), variability (V), and recommended follow-up intervals (FUT), supporting proactive, personalized measurement-based care.

The study demonstrates that continuous-time stochastic models can learn and update individual-level trajectories of suicidal ideation from irregularly sampled clinical data, aligning with the needs of measurement-based care. The superior performance of an individual-specific Wiener process suggests that short-term variability dominates trajectories in this cohort, with no universal pull toward a single long-term level. However, some individuals appear to have stable points, implying that mixture models or flexible structures accommodating both random-walk and mean-reverting dynamics could further improve personalization. By quantifying IHIP, V, and FUT, clinicians can gauge near-term risk, volatility, and how quickly prior information loses predictive value, informing tailored monitoring schedules and interventions. The observed heterogeneity in diffusion underscores that variability lies on a continuum rather than discrete categories; thus, individualized modeling is essential. Monitoring frequency should be higher when uncertainty is greatest (baseline or high variability) and can be reduced as stability is learned. Using individual-level predictions as a reference can help flag unexpected changes when new observations fall outside predicted ETIs, enabling proactive adjustments in care.

This work introduces and validates a continuous-time, hierarchical Bayesian framework to learn and dynamically update individual trajectories of suicidal ideation from naturalistic clinical data, producing personalized summaries (IHIP, V, FUT) to support measurement-based care. Key contributions include handling irregular observation times, mapping latent states to bounded clinical scales, and providing clinically interpretable monitoring recommendations based on information decay. Future research should: (1) incorporate abrupt state transitions and time-varying parameters; (2) explore mixture models that capture both random-walk and mean-reverting dynamics; (3) extend to multiple routinely monitored domains; (4) integrate the modeling pipeline into digital platforms for real-time clinical decision support; and (5) employ computationally feasible cross-validation strategies for model assessment when possible.

The sample comprises individuals engaged with services using the Innowell platform, potentially limiting generalizability. Data may be missing not at random due to optional data entry, introducing bias in observed ideation patterns. The current model does not capture the binary aspect and abrupt onset of ideation (e.g., transitions from zero to non-zero), nor dynamic changes in diffusion; baseline covariates were not explicitly modeled. Model comparison relied on WAIC rather than cross-validation due to computational constraints. For illustrative dynamic updating, population parameter uncertainty was fixed (using MAP estimates), not propagated into individual updates, potentially understating predictive uncertainty.