Artificial Intelligence and Precision Psychology for Predictive Mental Health: Integrating Climate Stress, Environmental Exposure, and Personalized Psychological Interventions

Abstract

The evidence base for effects of climate change and environmental exposure on mental health is virtually limited to average effects across groups. To quantify within-person associations between short-term heat and particulate exposure and daily psychological distress and their lag structure, as well as to identify moderators and one candidate mediator; To test, in silico, the efficacy of risk-triggered delivery of a brief psychological module using a fixed intervention budget compared to universal and random delivery; To determine the incremental value of environmental exposure features over clinical and digital-phenotyping features for 14-day symptom deterioration, including the lag structure of these features; To evaluate, in silico, the efficacy of risk-triggered delivery of a brief psychological module using a fixed intervention budget, compared to universal and random delivery. We created a synthetic population of 4,000 adults distributed in 5 climate regimes, experienced over an 180-day period (720,000 person-days). Published meta-analytic estimates of the effects of heat, particulate matter and greenness were used as anchors for structural parameters of the data-generating process. The within-person associations were estimated with person-mean centred regression and confirmed with random-intercept mixed models, with cluster-robust standard errors. Four prediction models were created (penalised logistic regression using clinical variables; gradient boosting using digital phenotyping; gradient boosting with environmental exposures added; multilayer perceptrons using flattened daily sequences). These models were evaluated using participant-grouped five-fold cross-validation and leave-one-site-out internal–external validations. Four delivery policies were tested within a framework of an assumed intervention effect based on the meta-analytic estimate of the JIT adaptive interventions intervention effect. Cumulative heat excess (lag 0–3) was associated with higher daily distress (β = 0.51 points per °C, 95% CI 0.47–0.55, p < .001; d = 0.079) as was cumulative PM2.5 (lag 0–6; β = 0.72 per 10 µg m−3, 95% CI 0.68–0.76, p < .001; d = 0.112). The impacts were greatest at lag 1596-1629 days for heat and lag 1596-1629 days for particulates. Trait climate anxiety amplified, and residential greenness and cooling access attenuated, the heat–distress slope. Sleep duration mediated 24.2% of the total heat effect (indirect β = 0.117, 95% CI 0.110–0.125). For 14-day deterioration (4,420 events among 44,000 person-windows; 10.0%), AUROC rose from 0.744 (95% CI 0.737–0.752) for clinical variables alone to 0.855 (0.850–0.860) with digital phenotyping, but adding the full environmental block yielded only 0.863 (0.858–0.868), an increment of 0.008 (95% CI 0.006–0.009). The results of the block-wise permutation analysis indicated that heat and air-quality features accounted for negligible amounts of discrimination (ΔAUROC 0.001 and 0.003 respectively) and that residential greenness, a stable person-level attribute, accounted for 0.015. The calibration slopes were approximately unity and performance was stable across sites (0.853–0.868) and subgroups (0.854–0.872). The delivery at 30% budget trial resulted in an absolute risk reduction of 1.29 percentage points (95% CI 1.18–1.39) compared with the number of prompts per event averted (23.3) at random allocation at the same budget (0.53, 95% CI 0.47–0.60) and universal delivery (1.77, 1.64–1.88 prompts per event averted). The individual risk prediction is a complementary approach to environmental exposure. Within-person associations of short-term exposure to heat and short-term exposure to particulate matter with distress are reproducible, lagged and moderated, and are not particularly helpful for discriminating who will deteriorate; these are important for understanding mechanism and timing delivery