Suicide remains a significant public health problem despite advancements in mental health treatment. Social media data offers a promising avenue for longitudinal risk assessment, as individuals may disclose suicidal thoughts and risk factors online more readily than to clinicians. Previous research highlights the potential of social media data analysis to identify suicidal ideation, but most studies focus on tweets explicitly mentioning suicide, neglecting subtle indicators of distress predictive of suicide risk. This study aimed to create a machine learning model that predicts individual-level future suicidal risk from social media data *before* any mention of suicidal thoughts, and validate this model at a population level.

Existing research exploring social media data's role in suicide risk assessment is limited. While several studies have analyzed Twitter data for explicit mentions of suicidal thoughts or attempts, few have considered subtle indicators of distress. Most studies lack the capability to predict suicidal ideation before its occurrence or to model temporal risk. One study attempted temporal risk modeling using latent suicide topic analysis, but its reliance on prior mentions of suicide limited its applicability to individuals without a history of expressing suicidal thoughts. This research addresses this gap by using machine learning algorithms integrating psychological theories of suicide, such as the interpersonal psychological theory of suicide (IPTS), the hopelessness model, and the associations of depression, anxiety, and insomnia with suicide risk, to predict risk before explicit mentions of suicidal ideation and to model temporal risk.

The study used a large dataset of tweets (512,526 from 283 suicidal ideation (SI) cases and 3,518,494 from 2655 controls). Neural networks were trained to infer psychological weights from text for constructs like stress, loneliness, burdensomeness, hopelessness, depression, anxiety, and insomnia. The performance of these neural networks was validated against binary adaptations of psychometrically validated scales. A random forest model was then trained using neural network outputs and sentiment polarity metrics to predict binary SI status. Bootstrap aggregating addressed class imbalance. The model's performance was evaluated using several metrics, including AUC, sensitivity, specificity, and positive predictive value. Temporal prediction was assessed using a sliding window approach and analyzing the frequency of model scores above an individual-specific threshold. Regional validation involved correlating algorithm scores from randomly sampled tweets within US counties with county-wide suicide death rates obtained from the Centers for Disease Control and Prevention (CDC). Age and sex estimations were performed using the M3Inference tool. Statistical analyses included AUC calculations, logistic regression, and Kendall's tau correlation.

The SAIPH model accurately predicted suicidal ideation events with an AUC of 0.88 (95% CI 0.86–0.90). The AUC was higher (0.90) for individuals with recurrent suicidal ideation and lower (0.80) for those with a single SI event. The model also distinguished individuals with a history of suicide attempts or plans from those without, achieving an AUC of 0.75. Model performance did not vary significantly by sex, but was stronger for younger individuals. Temporal analysis revealed that peak occurrences of scores above an individual-specific threshold indicated a ~7-fold increased risk of SI within the next 10 days (OR = 6.7 ± 1.1, P = 9 × 10⁻⁷¹). Regional validation demonstrated a significant correlation between aggregated SI scores and county-wide suicide death rates, especially in younger age groups (15–44 years). The study found a significant association of mean model scores with the time to suicide in celebrity suicide cases. Analyzing the time to suicide from peak frequency scores revealed a significant peak approximately 20 days before death. A minimum of 8-16 days of data were necessary to predict county-specific death rates, with the discrepancy potentially attributed to national events, such as mass shootings.

The SAIPH model demonstrates the potential of using machine learning to predict suicidal ideation from social media data, offering a novel approach to suicide risk assessment. The model's ability to predict both risk and temporal patterns highlights its potential as a clinical decision-making tool. The findings support previous research linking social media activity to suicidal behavior. The model's superior performance in predicting SI in younger individuals aligns with Twitter usage patterns. The regional validation further strengthens the model's relevance to real-world suicide rates, suggesting potential applications in monitoring suicide prevention or postvention efforts. The model's sensitivity to major events, like mass shootings, suggests its ability to capture immediate shifts in population-level distress. Future research could explore applications in smaller geographical areas, evaluating the impact of interventions, and addressing limitations related to data representation of various demographics.

This study successfully developed and validated SAIPH, a machine learning algorithm capable of predicting future suicidal ideation risk from social media data. The model's ability to predict both individual risk and temporal patterns represents a significant advancement in suicide risk assessment. Further research should focus on refining the algorithm, broadening its applicability to diverse populations, and exploring its integration into clinical practice to optimize suicide prevention efforts.

The study's limitations include the reliance on Twitter data, which may not capture the full spectrum of suicidal thoughts and behaviors. The age and sex estimations were based on an algorithm and could introduce inaccuracies. The study's focus on specific psychological theories might limit its ability to capture other risk factors, such as substance abuse. The control group selection strategy could be improved for more robust analysis. Lastly, the study lacked detailed demographic data on participants, limiting the understanding of its generalizability across different populations.