Publications by Year: 2024

Psychiatric trials have some of the lowest success rates across therapeutic areas, resulting in decreased investment in psychopharmacological drug development even as the need for more effective treatments grows. Digital measures and digital biomarkers (DBMs) provide one potential avenue for ameliorating three of the largest problems impeding clinical trial success in psychiatry: diagnostic heterogeneity, endpoint subjectivity, and high placebo response rates. First, DBMs may address heterogeneity and comorbidity in psychiatric nosology by identifying predictive DBMs of treatment response via the targeting of drugs to psychiatric subtypes. Second, DBMs can provide objective measures of physiology and behavior that when grounded in meaningful aspects of health (MAH) could support use for regulatory decision-making. By objectively and continuously measuring aspects of a patient's disease that the patient wants to improve or prevent from getting worse, DBMs might provide clinical trial endpoints that are more sensitive to treatment effects as compared to traditional clinician-reported outcomes. Lastly, DBMs could help address challenges surrounding high placebo response rates. Development of predictive DBMs of placebo response may allow for improved enrichment study designs to reduce placebo response. Objective digital measures may also be more robust against the placebo effect and offer an improved study endpoint alternative. Successful deployment of DBMs to address the historical challenges facing psychiatric drug trials will require close collaboration between industry, academic, and regulatory partners.

As the global prevalence of exposure to traumatic events rises, there is a growing need for accessible and scalable treatments for trauma-related disorders like posttraumatic stress disorder (PTSD). Intrusive reexperiencing symptoms, such as trauma-related intrusive memories (TR-IMs), are central to PTSD and a target of gold-standard treatments that are effective but resource-intensive. This study examined whether completing a brief ecological momentary assessment (EMA) protocol assessing the occurrence and phenomenology of TR-IMs was associated with decreases in intrusion symptom severity. Trauma-exposed adults (N = 139) experiencing at least 2 TR-IMs per week related to a DSM-5 criterion A traumatic event completed a 2-week EMA protocol. During this period, they reported on 18 sensory-perceptual and affective qualities of their TR-IMs three times per day and on posttraumatic stress symptom severity at day's end. Longitudinal symptom measurements were entered into linear mixed-effects models to test the effect of Time on symptom severity. Over the 2-week protocol, intrusion symptom severity decreased, while other symptom cluster scores did not change. Within the intrusion symptoms, this effect was specific to TR-IMs and emotional reactivity to trauma reminders, and was not moderated by survey completion rate, total PTSD symptom severity, ongoing treatment, or time since trauma. This study was quasi-experimental and lacked a control group, therefore no definitive conclusions about clinical utility can be made. Nonetheless, these findings provide preliminary proof-of-principle and warrant future clinical trials assessing the clinical efficacy of EMAs of intrusive trauma memories as a scalable treatment option targeting intrusive memory symptoms.

Social isolation and social impairment are hallmarks of progression as well as predictors of relapse in psychiatric disorders. We conducted a pilot study to assess the feasibility of sensing the social activity phenotype and loneliness using active and passive markers collected using a smartphone application. The study included 9 schizophrenia and bipolar disorder patients followed in the Bipolar Longitudinal study for at least 1 month and for whom mobile communication data was collected using the Beiwe smartphone application. Subjects completed daily surveys on digital and in-person social activity, and feelings of being outgoing or lonely. We described the level and variability of social activity features. We employed k-means clustering to identify "important contacts". Further, we investigated whether social network-derived features of mobile communication are independent predictors of weekly counts of outgoing calls and text, weekly average self-reported digital social activity, and loneliness using mixed effect models and clustering with dynamic time warping distance. Subjects were followed between 5 and 208 weeks (number of days of observation = 2538). The k-means cluster analysis approach identified the number of "important contacts" among close friends and family members as reported in clinical interviews. The cluster analysis and longitudinal regression analysis indicate that the number of individuals a person communicates with on their phone is an independent predictor of perceived loneliness, with stronger evidence when "important contacts" only are included. This study provides preliminary evidence that the number of "important contacts" a person communicates with on their phone is a promising marker to capture subjects' engagement in mobile communication activity and perceived loneliness.

Technology is playing an increasing role in healthcare, especially in mental health. Traditional mental healthcare, whether in-person or via telehealth, cannot by itself address the massive need for services. Standalone technology such as smartphone apps, while easily accessible, have seen limited engagement and efficacy on their own. Hybrid care - the combination of synchronous in-person or telehealth appointments with the use of asynchronous digital tools such as smartphone applications, wearable devices, or digital therapeutics - has the potential to offer the best of both worlds, providing both increased access and higher engagement and efficacy. In this paper, we present a framework highlighting the key components of hybrid care models: digital intervention, human support, and target population. This framework can be used to evaluate existing models in the literature and in practice, identify areas of need and opportunity, and serve as a blueprint for key elements to consider when designing new hybrid care models.

Stress produces profound effects on behavior, including persistent alterations in sleep patterns. Here we examined the effects of two prototypical stress peptides, pituitary adenylate cyclase-activating polypeptide (PACAP) and corticotropin-releasing factor (CRF), on sleep architecture and other translationally-relevant endpoints. Male and female mice were implanted with subcutaneous transmitters enabling continuous measurement of electroencephalography (EEG) and electromyography (EMG), as well as body temperature and locomotor activity, without tethering that restricts free movement, body posture, or head orientation during sleep. At baseline, females spent more time awake (AW) and less time in slow wave sleep (SWS) than males. Mice then received intracerebral infusions of PACAP or CRF at doses producing equivalent increases in anxiety-like behavior. The effects of PACAP on sleep architecture were similar in both sexes and resembled those reported in male mice after chronic stress exposure. Compared to vehicle infusions, PACAP infusions decreased time in AW, increased time in SWS, and increased rapid eye movement sleep (REM) time and bouts on the day following treatment. In addition, PACAP effects on REM time remained detectable a week after treatment. PACAP infusions also reduced body temperature and locomotor activity. Under the same experimental conditions, CRF infusions had minimal effects on sleep architecture in either sex, causing only transient increases in SWS during the dark phase, with no effects on temperature or activity. These findings suggest that PACAP and CRF have fundamentally different effects on sleep-related metrics and provide new insights into the mechanisms by which stress disrupts sleep.

Artificial intelligence (AI)-based computational tools for deriving digital behavioral markers are increasingly able to automatically detect clinically relevant patterns in mood and behavior through algorithmic analysis of continuously and passively collected data. The integration of these technologies into clinical care is imminent, most notably in clinical psychology and psychiatry but also other disciplines (e.g., cardiology, neurology, neurosurgery, pain management). Meanwhile, ethical guidelines for implementation are lacking, as are insights into what patients and caregivers want and need to know about these technologies to ensure acceptability and informed consent. In this work, we present qualitative findings from interviews with 40 adolescent patients and their caregivers examining ethical and practical considerations for translating these technologies into clinical care. We observed seven key domains (in order of salience) in stakeholders' informational needs: (1) clinical utility and value; (2) evidence, explainability, evaluation and contestation; (3) accuracy and trustworthiness; (4) data security, privacy, and misuse; (5) patient consent, control, and autonomy; (6) physician-patient relationship; and (7) patient safety, well-being, and dignity. Drawing from these themes, we provide a checklist of questions, as well as suggestions and key challenges, to help researchers and practitioners respond to what stakeholders want to know when integrating these technologies into clinical care and research. Our findings inform participatory approaches to co-designing treatment roadmaps for using these AI-based tools for enhanced patient engagement, acceptability and informed consent.

Medical knowledge is context-dependent and requires consistent reasoning across various natural language expressions of semantically equivalent phrases. This is particularly crucial for drug names, where patients often use brand names like Advil or Tylenol instead of their generic equivalents. To study this, we create a new robustness dataset, RABBITS, to evaluate performance differences on medical benchmarks after swapping brand and generic drug names using physician expert annotations. We assess both open-source and API-based LLMs on MedQA and MedMCQA, revealing a consistent performance drop ranging from 1-10%. Furthermore, we identify a potential source of this fragility as the contamination of test data in widely used pre-training datasets.

The introduction of Large Language Models (LLMs) has advanced data representation and analysis, bringing significant progress in their use for medical questions and answering. Despite these advancements, integrating tabular data, especially numerical data pivotal in clinical contexts, into LLM paradigms has not been thoroughly explored. In this study, we examine the effectiveness of vector representations from last hidden states of LLMs for medical diagnostics and prognostics using electronic health record (EHR) data. We compare the performance of these embeddings with that of raw numerical EHR data when used as feature inputs to traditional machine learning (ML) algorithms that excel at tabular data learning, such as eXtreme Gradient Boosting. We focus on instruction-tuned LLMs in a zero-shot setting to represent abnormal physiological data and evaluating their utilities as feature extractors to enhance ML classifiers for predicting diagnoses, length of stay, and mortality. Furthermore, we examine prompt engineering techniques on zero-shot and few-shot LLM embeddings to measure their impact comprehensively. Although findings suggest the raw data features still prevail in medical ML tasks, zero-shot LLM embeddings demonstrate competitive results, suggesting a promising avenue for future research in medical applications.

Although resilience is a dynamic process of recovery after trauma, in most studies it is conceptualized as the absence of specific psychopathology following trauma. Here, using the emergency department AURORA study (n = 1,835 with 63% women), we took a longitudinal, dynamic and transdiagnostic approach to define a static resilience (r) factor, which could explain greater than 50% of variance in mental well-being 6 months following trauma and a dynamic resilience factor, which represented recovery from initial symptoms. We then assessed its neurobiological profile across threat, inhibition and reward processes using functional magnetic resonance imaging collected 2 weeks post-trauma (n = 260). Our whole-brain and study-wide Bonferroni-corrected results suggest that resilience is promoted by activation of regions involved in higher-level cognitive functioning, reward valuation and salience detection in response to reward, whereas resilience is hampered by posterior default mode network activation to threat and reward. These findings serve to generate new hypotheses for brain mechanisms that could promote dynamic and multifaceted components of resilience following trauma.

BACKGROUND: Ankle fracture treatment is predicated on minimal displacement, leading to abnormal joint contact area. The purpose of this investigation is to determine whether computed tomography (CT) detects subtle mortise malalignment undetectable by x-ray in supination-external rotation-II (SER-II) injuries.

METHODS: A total of 24 patients with SER-II injuries, as demonstrated by negative gravity stress radiography, were included. Medial clear space (MCS) measurements were performed on bilateral ankle x-rays (injured and contralateral, uninjured side) at several time points as well as bilateral non-weight-bearing CT performed once clinical and radiographic healing was demonstrated (mean = 66 days post injury, range = 61-105 days). Statistical analyses examined differences in measurements between both sides.

RESULTS: Final x-rays demonstrated no differences between normal and injured ankle MCS (P = .441). However, CT coronal/axial MCS measurements were different (P < .05). CT coronal MCS measured wider by a mean difference of 0.67 mm (P < .001).

CONCLUSION: There is a high incidence of subtle mortise malalignment in SER-II ankle fractures, as demonstrated by CT, which is undetectable when assessed by plain radiographs. Although clinical outcomes are yet unknown, there are important implications of the finding of confirmed, subtle mortise malalignment in SER-II injuries and the limitations of x-ray to detect it.

LEVEL OF EVIDENCE: Level III.