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

Routine outcome monitoring (ROM) in mental healthcare, facilitated by digital tools, has increased the availability of longitudinal patient data. This allows for a more personalized and objective understanding of an individual's mental health status, treatment response, and progress. Measurement-based care (MBC) leverages this data to inform clinical decisions and tailor interventions to individual needs, leading to improved outcomes and communication. However, clinicians often struggle with accurate prediction of patient outcomes. While predictive models exist, they are often limited in their ability to incorporate the ongoing and dynamic nature of MBC. Traditional models often predict outcomes over fixed periods without dynamically updating with new observations. The paper argues for continuous-time modeling as a better alternative for analyzing and predicting suicidal ideation trajectories, addressing challenges like irregularly spaced observations and missing data. Existing categorical prediction models for suicidal thoughts and behaviors have shown limited success. This study aims to improve prediction by utilizing higher-frequency longitudinal observations and a continuous-time modeling framework. Continuous time models are attractive because they can handle irregularly spaced intervals and missing data, be framed hierarchically to learn individual-level parameters constrained by population-level data, and incorporate Bayesian frameworks to account for uncertainty. This study investigates the use of continuous-time models (specifically Wiener and Ornstein-Uhlenbeck processes) for naturally collected clinical data on suicidal ideation, aiming to provide insights that can aid clinical decision-making.

The literature highlights the growing importance of ROM and MBC in improving mental health care. Studies have shown the benefits of MBC, including improved outcomes, retention, and clinician-patient communication. However, the literature also reveals the limitations of clinicians' predictive abilities and the shortcomings of existing predictive models in the context of MBC. Existing predictive models for suicidal ideation often perform poorly, with categorical predictions showing only slight improvement over chance. While several factors have been associated with suicidal ideation, these associations have not translated into high predictive values. The literature emphasizes the need for higher-frequency longitudinal observations to capture the fluctuations in suicidal ideation, and continuous time models are proposed as a suitable solution due to their ability to handle irregularly spaced data and missing values. While continuous time models have been applied in other fields, their application in mental health remains limited.

The study used data from 585 individuals (72.6% female, mean age 24.2 years) collected via the Innowell digital platform. The data included longitudinal measurements of suicidal ideation (SIDAS), along with other relevant mental health measures and demographic information. The researchers compared several continuous-time model parameterizations, including Wiener and Ornstein-Uhlenbeck processes, using a Bayesian framework and a particle filter method to account for the link between the unbounded continuous model space and the bounded, discretized observational scale. Model comparison was performed using the Widely Applicable Information Criterion (WAIC). The best-performing model was used to generate individual-level predictions and visualize dynamic updates as new observations were collected. The analysis included calculations of the probability of being in a high-ideation category (IHIP), future observational variability (V), and a recommended follow-up time (FUT). Correlations between the diffusion parameter and baseline characteristics were examined.

The Wiener process model with random effects for all parameters was identified as the best-performing model. This suggests that the typical trajectory of suicidal ideation is dominated by variability, with no long-term constant. The model yielded individual-level parameters representing baseline suicidal ideation and diffusion over time. The analysis revealed no clear relationship between these two parameters, though a significant portion of individuals showed zero suicidal ideation at baseline. The diffusion parameter, representing variability, correlated positively with recent suicide attempts and borderline correlated with mania-like experiences. Illustrative examples showcased individual trajectories with varying patterns (improvement, deterioration, moderate ideation, high variability, and stable low ideation). The dynamic prediction updates showed decreasing uncertainty in trajectory prediction as more observations were collected. The analysis provided IHIP, V, and FUT for each individual, offering quantifiable measures of suicidal ideation level and variability. The FUT values suggest more frequent monitoring is needed in the early stages of care and during periods of high variability.

The study's findings highlight the heterogeneity of suicidal ideation trajectories. The population was divided into those with zero suicidal ideation and those with diverse trajectories characterized by considerable uncertainty at baseline. This underscores the need for personalized and dynamic approaches to monitoring and intervention. The continuous-time models allow for the learning and updating of predictions as more data becomes available. This contrasts with traditional methods that rely on matching baseline observations to population averages. The study emphasizes the importance of individualized diffusion parameters in modeling suicidal ideation, suggesting a continuum of variability rather than distinct categories. The variability of suicidal ideation is linked to different potential outcomes and interventional approaches. The FUT provides a data-driven suggestion for optimized monitoring frequency, highlighting the need for more frequent monitoring during early stages of care and periods of high variability. The results suggest a proactive, rather than reactive, approach to managing suicidal ideation.

This study demonstrates the value of continuous-time trajectory models for personalizing the monitoring and management of suicidal ideation. The dynamic model provides valuable insights into individual trajectories, allowing for more informed clinical decisions. Future work should focus on refining the model by incorporating abrupt state changes and time-varying parameters, and on developing a user-friendly digital implementation for clinical use.

The study's limitations include the sample being restricted to individuals using the Innowell platform, potentially introducing biases in data collection. The model may not fully capture the binary aspect of suicidal ideation or other dynamic changes. The model comparison used WAIC instead of cross-validation, a potentially less robust method but necessary given computational constraints. The dynamic prediction updates made simplifications by fixing prior parameters using maximum a posteriori estimates.