Page 62 - GTM-4-2
P. 62

Global Translational Medicine                                        Sleep and emotion rhythmicity in tweets



            text-mining algorithms could be applied benevolently   method. Large-scale studies are required to assess the
            to identify individuals susceptible to emotional distress   stability of intra-individual interactions between dRARs
            from circadian and sleep-related disruption, which could   and emotional tweet content across a larger and diverse
            facilitate  the  administration  of  “Just-In-Time  Adaptive   population.
            Interventions,”  or signpost to appropriate support. One   From a translational perspective, future studies may
                        6
            of the reasons that existing suicide prevention approaches   consider assessment of clinical validity among known
            have had limited impact on suicide rates is that prevention   “at-risk” populations as all the lexicons used in NLP
            interventions are too slow for the dynamic temporal   methods such as sentiment analysis and LIWC are based on
            nature of suicide ideation, which fluctuates in real-  healthy individuals. Once the goal of these methodologies
            time. In particular, during the night, when suicide risk   becomes the identification of clinically salient emotional
            is substantially elevated and support mechanisms are   distress, one must consider the extent to which we
            scare, social media monitoring could provide timely and   can consider these normative datasets to be valid. We
            ecological support in the form of counseling, or referral   recommend the development of lexicons based on distinct
            to 24-h telehealth services when it is most needed. These   clinical populations (e.g., major depressive disorder,
            interventions may complement or even integrate with   generalized anxiety disorder), which may maximize the
            automated harm-prevention mechanisms present on    sensitivity of these models. Moreover, language by nature
            social media sites, such as “pop-ups” offering support   differs drastically among cultures, and societies. Therefore,
            resources whenever a user’s activity is flagged due to the   it is important to consider whether models trained using
            use of keywords associated with suicide. Nevertheless,   one particular language be transferred to another, as well
            the implementation of these methods presents a number   as whether these models perform equally well across
            of methodological and legal challenges, which we discuss   different populations. Equally important is the question of
            over the proceeding sections.
                                                               how differences in language are perceived across different
            3. Methodological challenges                       genders,  ethnic  identities,  races,  and  cultures,  and  how
                                                               their relationship with the 24-h cycle should be considered.
            3.1. Determining validation and alignment with     Given that certain cultures, ethnicities, and genders
            circadian rhythms                                  communicate emotion much more readily than others, it

            Much in the same manner that wearable devices which   is critical to consider how conspicuous language should be
            track sleep must be compared against gold-standard   before it is considered to be indicative of emotional distress.
            metrics of sleep, examinations of how sleep estimates from   Nevertheless, obtaining this demographic information
            dRARs align with predetermined benchmarks for accuracy   from social media may prove unreliable, due to the limited
            must be conducted, by assessing their synchronicity with   nature of the source data (a user’s profile). Much in the
            well-defined biological (e.g., circadian), environmental   same  way that disorder-specific  models  may  contribute
            (e.g., night/day or light/dark) or social rest-activity (e.g.,   toward increased precision for the detection of emotional
            work-leisure)  rhythms  when  measured  using  techniques   distress, models which adapt to the differences in sleep
            commonly held as accurate assays of these rhythms. For   behaviors (e.g., chronotype, bi-phasic sleeping or siesta
            instance, determination of the peak of dRARs activity   cultures) may prove beneficial. However, consideration
            relative to one’s dim light melatonin onset peak could   of  how to  considerately  integrate these  data with other
            provide a window into alignment with endogenous    sources of demographic meta-data must first be addressed.
            biological rhythms. Most relevant are comparisons with
            24-h waist-actigraphy (a gold-standard for rest-activity   3.2. Defining sleep window with minimal datapoints
            rhythms [RAR] assessments), which would represent a   Perhaps the most salient methodological challenge
            logical next-step in this validation pipeline and necessitate   associated with deriving RARs from social media use is
            comparison of how accurately measurements of L5 and   the continuity and completeness of the data. Thorough
            M10 (most active 10/24  h) align between dRARs and   sensitivity analyses to determine both the minimum
            movement RARs. Such validation studies should employ   amount of data are required to accurately determine
            standardized approaches and recommended guidelines   rhythms, as well as the precision points at which further
            (for further details see Smith et al. [2018] ) for traditional   data brings no further gain in model accuracy. These
                                             7
            RAR  assessment  in  performing  these  comparisons.   parameters should be determined to ensure the credibility
            Crucially, as we demonstrated this approach on a single   of predictions, as well as protection against unnecessary
            user, generalizability of these findings is limited, and   exposure of data when it is not necessary to inform models.
            should only be considered as a proof-of-concept of the   Although the present study focused exclusively on social



            Volume 4 Issue 2 (2025)                         54                              doi: 10.36922/gtm.5176
   57   58   59   60   61   62   63   64   65   66   67