Idiographic symptom networks fail to predict future symptom states across four ecological momentary assessment datasets

A central ambition of computational psychiatry is to model psychiatric symptom dynamics at the level of the individual. Symptom networks estimated from longitudinal ecological momentary assessment (EMA) recordings have emerged as the predominant framework for this goal, promising idiographic insights into the temporal structure of psychopathology and its treatment.

Yet whether these networks can be robustly estimated from the short, noisy, and irregularly sampled time series typical of EMA studies, and whether these models capture meaningful symptom dynamics at all, remains an open question.

We address this through the lens of predictive validity, using a rolling-origin out-of-sample forecasting scheme recently established as a benchmark for EMA forecasting [1]. Models are trained on all observations up to a given time point and evaluated on their ability to forecast the next observed symptom states.

In a systematic analysis spanning four EMA datasets covering different symptoms, sampling schemes, and estimation strategies, we compare the predictive performance of a range of symptom network models against simpler baselines. As network models, we consider the Vector Autoregressive (VAR) model — the predominant tool for network estimation — estimated via unregularized OLS, structural equation modelling, LASSO regularisation, and multilevel hierarchical modelling, as well as Almost-Linear Recurrent Neural Networks [2], a recent machine learning-based state space model capable of capturing nonlinear symptom dynamics. As baselines, we use the AR(1) model (a restricted VAR without cross-symptom connections) and two naive predictors: persistence (predicting the next step as the current value) and the time series median.

We find that in an overwhelming majority of participants (295/314), VAR-based symptom networks are outperformed by the simpler AR models, indicating that estimated cross-symptom connections actively worsen out-of-sample predictions. While VAR performance improves monotonically with training set size, it remains below AR performance in all datasets, and this improvement is accompanied by a progressive pruning of cross-symptom connections, suggesting that additional data eliminates spurious connections rather than uncovering true ones.

These results hold across all estimation methods and model classes, including nonlinear state-space models. Indeed, forecasting performance across all models is bracketed by two naive baselines: persistence frequently outperforms even the AR model at the next-step horizon, while the time series median proves hard to beat for day-ahead forecasts.

Complementary simulations confirm that under realistic EMA conditions, even when data are generated by a ground-truth VAR process with realistic network strengths and sampling rates, the underlying cross-symptom structure cannot be recovered with sufficient precision to yield a predictive advantage.

Taken together, our results demonstrate that cross-symptom connections remain unidentifiable under the conditions characteristic of real EMA data, calling into question the interpretive framework built on idiographic symptom networks and pointing toward concrete requirements in study design, sampling density, and time series length that future research must address.

References

1. Fechtelpeter J, Rauschenberg C, Goetzl C, Hiller S, Wierzba E, Emonds N, et al. Computational network models for forecasting and control of mental health trajectories in digital applications. npj Digit Med. 2026;9:70.
2. Brenner M, Hemmer CJ, Monfared Z, Durstewitz D. Almost-Linear RNNs yield highly interpretable symbolic codes in dynamical systems reconstruction. Advances in Neural Information Processing Systems. 2024.