Skip to main content
World Psychiatry logoLink to World Psychiatry
. 2024 May 10;23(2):288–289. doi: 10.1002/wps.21209

The future of dynamic networks in research and clinical practice

Laura F Bringmann 1
PMCID: PMC11083907  PMID: 38727056

A problem attracts other problems 1 . An individual having one symptom (e.g., feelings of sadness) often will have more symptoms over time (e.g., sleep disturbances). The observation that symptoms generate other symptoms has led to the highly influential psychopathological network theory, in which the focus is on how mental disorders can be seen as networks of interacting symptoms 2 . In such psychopathological networks, symptoms are the nodes, and the connections between them are termed edges.

The network theory posits that individuals can have different symptom networks, which can also change within a single person over time. Not only are the symptoms in the network likely to be different between individuals; the connections between the symptoms are also likely to differ. Whereas for person A having sleep disturbances may trigger suicidal thoughts, this may not be the case for person B. Subsequently, a central notion stemming from network theory is that interventions should not only target symptoms, but also the connections between them, breaking the vicious cycle.

In clinical research and practice, networks for patients are often inferred from time series data based on symptom and/or emotion questionnaires that are administered several times a day over, for instance, a time period of several weeks. Networks based on such data have been called person‐specific dynamic networks. Even though the models used for inferring these dynamic networks are not new, what is new is the visualization of these models, which makes network modelling appealing and tangible for both research and clinical practice 3 .

What is often overlooked in the field is the need to bridge the gap from an insightful theory to an actual model. To get from a network theory to a model there are, broadly speaking, two steps that need to be made. The first step concerns the kind of data to be collected from the patient, and the second regards the modeling of these data in such a way that it does justice to the theory.

Thus, the first step to a dynamic network model is collecting and selecting the necessary data. In the network theory, the focus has been on symptoms and emotions, and thus most person‐specific dynamic networks almost exclusively include symptoms and/or emotions items. However, further behavioral, context and cognitive items could be included in the networks 4 . Emotions and symptoms indicate how a patient is doing, whereas behavioral, context and cognitive items can lead to insights on why the patient is feeling a certain way. This can help in targeting therapeutic interventions, for example, by indicating behaviors, social contacts or thoughts that lead to feelings of sadness. Additionally, these items should be formulated and personalized in terms of content in collaboration with the patient, which can result in a higher chance of capturing crucial aspects of his/her daily life, and better reflect his/her interests and experiences in general 5 .

Other sources than questionnaire data should also be considered, such as social media, passive sensing, and many other available kinds of time series data. Collection of passive sensing data (e.g., location tracking of mobile phones) can lead to reduction of the burden on the patient, as the data are gathered automatically. In addition, it can provide new therapeutically relevant insights complementing questionnaire data. For example, if a patient with social anxiety indicates in a questionnaire that she followed a lecture, it is not clear from this whether she had a breakthrough and actually went to campus, or watched the lecture online at home, but location tracking data can provide this complementary information.

After the source of data has been chosen, a model needs to be identified to infer the network. With regard to this second step, models based on vector autoregression are the most commonly used to infer these person‐specific networks. A vector autoregressive model allows to study, for instance, reciprocal predictive effects: does stress predict one's sadness at the next time point, the other way around, or are both stress and sadness predictive of each other?

Although this model does justice to the dynamic structure of data, it is too limited to capture all aspects of the network theory. In fact, it cannot deal with change in the network, whereas this is often one of the main elements of interest (e.g., did the mood of the patient improve and why?). Furthermore, different variables measured at different time scales (e.g., sleep quality vs. mood) cannot be modeled simultaneously, meaning that often factors such as sleep quality (measured once a day) are not taken into account into the network, even though they are known to affect one's mood (mood is often measured several times a day). Additionally, these models are very data hungry, meaning that many time points (more than 100 per individual) are often needed to model all relevant variables in the network 6 .

Beyond the many different choices that one could make in the pursuit of a network model, one could also take a different direction altogether. Network theory itself is not tied to any model, nor to a number of variables. Thus, fitting a vector autoregressive model on all relevant variables to infer a dynamic network is neither necessary nor per se advisable, as including more variables but not increasing the number of time points will only decrease the accuracy of the estimated network 7 .

Instead of the rather ambitious and broad goal of trying to unravel the dynamic processes between symptoms, a more viable approach could be to test specific and targeted questions, which often emerge naturally between patient and therapist (e.g., in functional analysis) 8 . Studying the relation between only two variables can already be of interest: for example, studying whether a specific physical activity leads to a positive change in one's sleep pattern. This is still in line with the network theory (how do variables influence each other), but, given the focus on change, should not be studied with the standard vector autoregressive model 9 . Rather, it could be studied with, for instance, regime‐switching models. With these models, one could investigate whether there are different regimes in the data, and whether the regime of more physical activity (vs. less physical activity) is associated with a positive change in one's sleep pattern 4 . In general, different questions might require different (network) models. Therefore, a broadening of our available arsenal of statistical techniques is also essential to optimize the move from network theory to an actual model.

To summarize, we see that until now dynamic networks have been implemented in a rather restrictive way, in which often only specific variables and one specific model are commonly used. Importantly, the network theory puts no constrains on the data and the model used to come to a person‐specific network. This means that in the future we do not have to hold on to specific data or models. In the end, researchers and clinicians alike should explicate how they want to apply the network theory to the specific patient and situation, and then consider what kind of data and model is best suited for that purpose.

REFERENCES


Articles from World Psychiatry are provided here courtesy of The World Psychiatric Association

RESOURCES