Using biobehavioral technologies to effectively advance research on negative symptoms

Negative symptoms have been a core component of schizophrenia since the pre‐neuroleptic era and are related to particularly poor clinical outcomes (e.g., in terms of recovery, quality of life, subjective well‐being). They are often one of the first markers of illness risk, emerging in the premorbid and prodromal phases, and in many patients remain stable throughout the first episode and chronic phases of illness1. Unfortunately, the mechanisms underlying these symptoms are poorly understood, and currently available treatments are unsatisfactory and palliative at best2.

To date, our understanding of negative symptoms is almost entirely dependent on psychometrically‐supported clinical rating scales. Within the last few decades, a literature has emerged using “biobehavioral” technologies to measure negative symptoms from objective vocal, language, facial, decision making, gestural, electrophysiological, neurobiological, and reaction time measures.

While clinical ratings reveal abnormalities on the scale of three to seven standard deviations in patients versus non‐psychiatric controls1, 3, group differences in biobehavioral measures tapping their underlying constructs are much smaller, if not altogether absent. For example, despite dramatic clinically‐rated patient abnormalities in alogia, blunted affect, anhedonia, and avolition1, 3, 4, studies of computerized speech analysis, hedonic experience, and motivation often find negligible or small effect size abnormalities5, 6.

Moreover, biobehavioral measures often show surprisingly modest and negligible correlations with conceptually overlapping negative symptom ratings6, 7, or similarly sized correlations to a wide array of non‐negative symptom ratings8. Though statistical significance may be reported in isolated studies for specific biobehavioral features, findings often do not replicate across studies, and the magnitude of effects are generally well below levels suggestive of acceptable convergent validity3.

In light of this surprisingly low convergence between clinical ratings and biobehavioral technologies, it is tempting to champion one of the two as being superior for measuring negative symptoms. Clinical ratings tend to be consistent across trained raters (i.e., reliability) and are associated with a wide range of important clinical variables (i.e., validity). On the other hand, biobehavioral technologies show near perfect reliability (assuming static recording conditions), and are instrumental to modern biometrics for measuring human functions that potentially underlie negative symptoms. So how can clinical ratings and biobehavioral technologies both be “reliable” and “valid” for measuring negative symptoms, yet show such surprisingly modest convergence?

Clinical ratings and biobehavioral technologies are fundamentally different in how they scale negative symptoms. Clinical ratings reflect the integration of an impressive number of complicated data streams over dynamic conditions. Using clinical rating scales, for example, a clinician is able to derive a gross ordinal value (e.g., “mild”) regarding blunted affect from a highly complex “spectrum” of vocal, verbal, facial and gestural data that fluctuate over time, questions and a changing environment.

Unfortunately, an individual negative symptom rating cannot be systematically “downscaled” for quantification. For example, deconstructing clinically‐rated blunted affect proves impossible in terms of which exact psychomotor channel was abnormal, when it was abnormal, or what factors may have mitigated the abnormality. In contrast, biobehavioral technologies afford the opportunity to precisely quantify continuous streams of highly specific speech, facial and gestural data, and to isolate, upscale, downscale and integrate them in a myriad of ways.

Unfortunately, it is unclear how to best do this. Which of the thousands of potential features computed from, for example, vocal analysis should be used, and how should they be weighted when they do not converge? Should computerized facial analysis reflect aggregate statistics during an entire interview, only when patients are speaking, during key temporal epochs, or following specific questions? In short, clinical ratings provide a view of the forest at the expense of being able to see the trees, whereas biobehavioral technologies afford the opposite.

Ideally, there would be a way to measure negative symptoms by marrying “low‐resolution” but “ecologically‐valid” ratings with “high‐resolution” but “dizzyingly‐complex” biobehavioral data. Within computational psychiatry more generally, an emerging “big data” literature now exists modeling various clinical diagnoses and ratings using biobehavioral features. Within these studies, models are typically built and optimized using a single biobehavioral channel without regard to other biobehavioral channels or to temporal, contextual or other dynamic factors.

While impressive accuracy rates are being reported, the models produced from this literature have yet to progress beyond “proof of concept” and seem particularly ill‐equipped for modeling negative symptoms. This is because clinician ratings are typically derived from multiple behavioral domains, and it is difficult to evaluate even one of these domains without considering context. For example, a patient's failure to activate his/her zygomaticus major muscle, language production, or reward systems can only be interpreted as abnormal when context is taken into account. After all, non‐patients are not actively smiling, talking or experiencing joy the vast majority of their day. Further complicating this issue is the reality that the behaviors underlying negative symptoms vary dramatically across and within people as a function of neurodevelopmental and cultural factors. In this manner, norms regarding smiling, talking and experiencing joy are very difficult to derive.

So, how can biobehavioral‐based modeling of negative symptoms progress? Technology and software systems have progressed so that they are affordable, reliable, highly sensitive, and unobtrusive, with a high potential for large‐scale international data collection across a broad range of behavioral domains. This allows for biobehavioral data collection that extends well beyond the relatively artificial confines of the clinic or research laboratory. For this, ecological momentary and ambulatory assessment methods, such as geolocation, passive vocal recording, activity tracking, and social media analysis, can complement existing measurement approaches.

Efforts to validate these technologies for understanding negative symptoms are currently underway. However, integrating and understanding these data within a network that can handle temporally and contextually dynamic data is a complex computational obstacle. Relatively simplistic “connectionist” and dynamic algorithms are being developed for many important human functions, and there is a growing field of understanding “networks of networks” to model complex interactions (e.g., “network medicine”)9.

In sum, existing clinical rating measures offer a level of precision that has not promoted advances in understanding underlying mechanisms and developing targeted treatments of negative symptoms. This reflects a “scalability” problem that can potentially be solved by modeling clinical ratings with multidimensional biobehavioral data streams.

Developing biobehavioral models can help pinpoint neurobiological and environmental mechanisms, modify them in real time using biobehavioral feedback, and develop, test and individualize targeted psychosocial and pharmacological agents to ameliorate their severity, and ideally, develop treatments.

Accurate modeling of negative symptoms is a complex endeavor, and an exciting computational opportunity that may advance multidisciplinary sciences and bring together researchers, patients and their support teams from around the world.