A Social Media Study on the Effects of Psychiatric Medication Use

May combine two claims.

Most psychiatric drugs are linked to a post-treatment increase in negative affect and a decrease in positive affect and cognition.

“Our results show that most drugs are linked to a post-treatment increase in negative affect and decrease in positive affect and cognition.”

May combine two claims.

SSRIs are associated with worsening symptoms, whereas TCAs lead to improvements.

“Clinically speaking, SSRIs are associated with worsening symptoms, whereas TCAs lead to improvements.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The distribution of top drugs in the dataset aligns with external surveys on the most prescribed psychiatric drugs, which found that the top 5 antidepressants captured over 70% of prescription volumes.

“This aligns with external surveys on the most prescribed psychiatric drugs in that time which found that the top 5 antidepressants captured over 70% of the prescription volumes (Scripts 2018; Gohol 2018).”

• cites Scripts 2018Not checked yet

  Not checked in the free run.

• cites Gohol 2018Not checked yet

  Not checked in the free run.

May combine two claims.

91% of Cohen's d values fall outside the small-difference range, suggesting the Treatment group significantly differed from the Control group.

“We find that 91% of our values fall outside this range, suggesting the Treatment significantly differed from the Control group.”

May combine two claims.

Most medications lead to similar directionality of effects on all outcomes; depression, anxiety, psychosis, and suicidal ideation increase for Treatment users in the After period of reported medication use.

“We observe many interesting patterns here, such as most medications lead to similar directionality of effects on all the outcomes, e.g., all of the outcomes, depression, anxiety, psychosis, and suicidal ideation increase for the Treatment users in the After period of reported medication use.”

May combine two claims.

TCAs had the largest improvement in these measures among drug families; roughly half of their users had higher positive affect and cognition.

“Among the drug families, we find that the TCAs show the greatest improvement in these measures, with about half of their users showing increased positive affect and cognition.”

May combine two claims.

In Figure 7, the affect and cognition outcome measures decreased mostly in lower-valued strata and increased in higher-valued strata.

“Next, Figure 7 shows that these outcome measures decrease mostly in the lower-valued strata and increase in the higher valued ones (Figure 7).”

May combine two claims.

An independent sample t-test reveals statistical significance in the differences (t in [-9.87, 10.96]; p<0.001), confirming that after self-reported medication use, Treatment users showed significant changes in outcomes.

“An independent sample t-test further reveals statistical significance in these differences (tϵ[−9.87,10.96]; p¡0.001), confirming that after the self-reported use of medications, the Treatment users showed significant changes in outcomes.”

May combine two claims.

Fluoxetine and Duloxetine were more effective for people who used more cognitive words, a group usually with less cognitive impairment.

“Similarly, Fluoxetine and Duloxetine shows better efficacy in those individuals with greater usage of cognitive words—typically those who show lower cognitive impairment”

The quote only partly states this claim.

In Sertraline models, first-person singular and auxiliary verb usage had negative coefficients linking ITEs to psycholinguistic attributes.

“For Sertraline, the use of first person singular and auxiliary verb shows negative coefficients”

May combine two claims.

TCAs and TeCAs had the strongest improving effects; most of their users were in strata where outcome measures decreased.

“From the perspective of drug families, the TCAs and the TeCAs show the greatest improving effects, with the majority of their users belonging to strata with decreased effects in the outcome measures.”

The quote only partly states this claim.

Sertraline, Escitalopram, and Fluoxetine-the most popular medications-were linked to worsening in these outcomes.

“These also include the most popular medications such as Sertraline, Escitalopram, and Fluoxetine.”

May combine two claims.

Most medications looked similar when aggregated, but their effects differed across strata.

“Although most medications show similar effects at an aggregated level, we find differences in their strata-wise effects distributions (Figure 7).”

May combine two claims.

Among drug families, SSRIs had the greatest worsening for depression, anxiety, psychosis, and suicidal ideation.

“All of these are classified as SSRIs—the family which shows the most worsening in these outcomes among the drug families.”

May combine two claims.

The study developed machine learning models to assess psychopathological states including mood, cognition, depression, anxiety, psychosis, and suicidal ideation, and these models were expert-validated.

“We first develop expert-validated machine learning models to assess psychopathological states known to be affected by psychiatric medications, including mood, cognition, depression, anxiety, psychosis, and suicidal ideation, as given in the literature (Coupland et al. 2011).”

• cites Coupland et al. 2011Not checked yet

  Not checked in the free run.

May combine two claims.

The findings reveal signals of the most common effects of psychiatric medications over a large population, with potential for improved characterization of their occurrence.

“Clinically, our findings reveal signals of the most common effects of the psychiatric medications over a large population, with the potential for improved characterization of their occurrence.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The classifiers were tuned using k-fold cross-validation with k=5.

“We use a balanced number of examples for the two classes in training, and we tune the parameters of the classifiers using k-fold (k=5) cross-validation (Chandrasekharan et al. 2018).”

• cites Chandrasekharan et al. 2018Not checked yet

  Not checked in the free run.

May combine two claims.

A major weakness of randomized controlled trials is that they are often conducted on individuals who may significantly differ from actual patients and are often not externally validated to a larger, more representative population.

“A major weakness of these trials is that they are often conducted on individuals who may significantly differ from actual patients, and often, they are not externally validated to a larger and a more representative population (Hannan 2008).”

• cites Hannan 2008Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Web search logs improve detection of adverse effects by 19% compared to an offline approach.

“White et al. (2016) found that web search logs improve detection of adverse effects by 19%, compared to an offline approach.”

• cites White et al. (2016)Not checked yet

  Not checked in the free run.

May combine two claims.

The classifiers show mean cross-validation accuracy between 0.79 and 0.88 and mean test accuracy between 0.81 and 0.91.

“These classifiers show a mean cross-validation accuracy ranging between 0.79 and 0.88 and mean test accuracy ranging between 0.81 and 0.91.”

May combine two claims.

The personal medication intake classifier uses an SVM model with a linear kernel and shows a mean k-fold cross-validation accuracy and F1-score of 0.82 each.

“The classifier uses an SVM model with linear kernel and shows a mean k-fold (k=5) cross-validation accuracy and F1-score of 0.82 each.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

TCAs have a broad spectrum of neurotransmitter affinities, which may often lead to undesirable side effects such as liver toxicity, excessive sleepiness, and sexual dysfunction.

“However, TCAs have a broad spectrum of neurotransmitter affinities, which may often lead to undesirable side effects, such as liver toxicity, excessive sleepiness, and sexual dysfunction (Frommer et al. 1987).”

• cites Frommer et al. 1987Not checked yet

  Not checked in the free run.

May combine two claims.

A linguistic equivalence test between the Reddit training dataset and the Twitter unseen dataset found a vector similarity of 0.95 for the top 500 n-grams, showing high content similarity and justifying the transfer approach.

“Using word-vectors (pre-trained on Google News dataset of over 100 billion tokens), we find the vector similarity of the top 500 n-grams in the Reddit and Twitter corpuses to be 0.95. This shows high content similarity across the two platforms, in turn justifying the transfer approach.”

May combine two claims.

The top features of the classifiers align with similar mental health classifiers built on Twitter, indicating construct validity of the transferred classifiers.

“Second, we find that the top features of these classifiers align with that of similar mental health classifiers built on Twitter to identify depression (De Choudhury et al. 2013), anxiety (Dutta et al. 2018), stress (Lin et al. 2014), psychosis (Birnbaum et al. 2017), and suicidal ideation (Burnap et al. 2015). This indicates the construct validity of the transferred classifiers.”

• cites De Choudhury et al. 2013Not checked yet

  Not checked in the free run.

• cites Dutta et al. 2018Not checked yet

  Not checked in the free run.

• cites Lin et al. 2014Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The study uses the well-validated Linguistic Inquiry and Word Count lexicon to quantify psycholinguistic shifts in affect and cognition.

“In particular, we use the changes in the normalized occurrences of words in these categories per the well-validated Linguistic Inquiry and Word Count (LIWC) lexicon (Tausczik and Pennebaker 2010).”

• cites Tausczik and Pennebaker 2010Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Increasing the selectivity of the target of action to individual neurotransmitters would, in theory, limit the incidence of side effects while maintaining the effectiveness of TCAs.

“Several other compounds have since been introduced whose development was guided by the idea that increasing the selectivity of the target of action to individual neurotransmitters would, in theory, limit the incidence of side effects while maintaining the effectiveness of TCAs (Lopez-Munoz and Alamo 2009).”

• cites Lopez-Munoz and Alamo 2009Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

These symptoms are comorbid with mood disorders.

“Interestingly, these symptoms are also comorbid with mood disorders (Rosenblat et al. 2016)”

• cites Rosenblat et al. 2016Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The clinician co-author established the validity and relevance of the final list of psychiatric drugs.

“Our clinician co-author established the validity and relevance of this final list.”

May combine two claims.

Clinical trials suffer from limitations such as non-standardized study design, confounding factors, and restrictive eligibility criteria.

“In addition, these clinical trials suffer from limitations such as non-standardized study design, confounding factors, and restrictive eligibility criteria (Lexchin et al. 2003).”

• cites Lexchin et al. 2003Not checked yet

  Not checked in the free run.

May combine two claims.

Clinical trials can be biased because they are conducted and funded by pharmaceutical companies, and they are rarely replicated in large populations.

“However, these trials can be biased, being conducted and funded by pharmaceutical companies, and are rarely replicated in large populations (Lexchin et al. 2003).”

• cites Lexchin et al. 2003Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Several top n-gram features such as depression, stress, hope, help, and feel are contextually related to mental health.

“Several top n-gram features such as depression, stress, hope, help, and feel, are contextually related to mental health.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The findings were corroborated with a co-author who is a board-certified psychiatrist.

“Our findings were corroborated with our co-author who is a board-certified psychiatrist.”

May combine two claims.

The treatment and placebo dates follow similar distributions, with a Kolmogorov-Smirnov statistic of 0.06 indicating similarity.

“We measure the similarity in their distribution using Kolmogorov–Smirnov test to obtain an extremely low statistic of 0.06, indicating similarity in the probability distribution of treatment and placebo dates (Figure 3b).”

May combine two claims.

Two researchers manually inspected 170 randomly selected Twitter posts and found high agreement of 87% between manual ratings and the classifiers' labels.

“Two researchers manually inspected 170 randomly selected Twitter posts on mental health symptoms, spanning both user samples. Using the methodology outlined in Bagroy et al. (2017) that draws up the DSM-5 clinical framework, they rated each Twitter post on a binary Likert scale (high/low) to assess levels of expressed depression, anxiety, stress, psychosis, or suicidal ideation. We find high (87%)”

• cites Bagroy et al. (2017)Not checked yet

  Not checked in the free run.

May combine two claims.

Digital pharmacovigilance studies, particularly those using social media, have largely targeted named adverse effects of drugs and have not measured broader forms of symptomatic changes longitudinally.

“Nevertheless, we observe a gap that digital pharmacovigilance studies, particularly those using social media, have largely targeted the named adverse effects of drugs (e.g., “headache”, “palpitations”, “nausea”), and have not measured broader forms of symptomatic changes longitudinally.”

Classified as probably not a substantial claim.

Even with a psychiatrist's corroboration, our view is limited to what an individual's social media data shows.

“Despite corroboration by a psychiatrist, we are limited by what can be observed from an individual’s social media data.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Adding this information to drug safety monitoring could supplement traditional sources and raise awareness of new safety problems.

“Including this information in technologies for drug safety monitoring can therefore complement traditional sources, and improve awareness regarding emerging safety issues in a spontaneous fashion”

May combine two claims.

Self-report bias complicates which effects we see, and we can only observe effects typical of the people in our data.

“Self-report bias further complicates the types of effects that we observed—different individuals respond differently, as shown in our results, however, our observations are limited to only the types of effects that characterize the individuals in our data.”

Classified as probably not a substantial claim.

We advise against generalizing our findings on psychiatric medication effects to whole populations.

“For these same reasons of sampling bias, we caution against drawing population-wide generalizations of the effects of psychiatric medication usage.”

Classified as probably not a substantial claim. May combine two claims.

Social media has no placebo concept, which trials use to remove the confound that merely perceiving treatment causes non-specific effects.

“First, we do not have the notion of a placebo, used to eliminate the confound that simply the perception of receiving a treatment produces non-specific effects.”

May combine two claims.

The paper offers a new source of collateral information for digital therapeutics and more personalized, evidence-based drug treatment.

“This paper provides a new source of collateral information to support digital therapeutics (Fisher and Appelbaum 2017) and enhance evidence-based, personalized pharmacological treatment.”

• cites Fisher and Appelbaum 2017Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Our symptom measures, like depression or suicidal ideation, require further clinical validation, such as against DSM-5 or RDoC.

“Further, the symptomatic outcomes themselves, such as measures of depression or suicidal ideation, need additional clinical validation, e.g., based on DSM-5 criteria (APA and others 2013), or the Research Domain Criteria (RDoC) introduced by the National Institutes of Mental Health (Insel et al. 2010).”

• cites APA and others 2013Not checked yet

  Not checked in the free run.

• cites Insel et al. 2010Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

For public health, our methods could enable technologies that detect early warnings of adverse effects from psychiatric drugs.

“From a public health perspective, our methods offer the potential to build technologies that surface early warning signs of adverse effects related to psychiatric drug use.”

May combine two claims.

The FDA issued two social media guidelines for prescription drug risk-benefit analysis, but they concern promotion and create no enforceable rights or duties.

“Although the FDA has released two guidelines on the use of social media for the risk-benefit analysis of prescription drugs (Sarker et al. 2015), they focus on product promotion and “do not establish legally enforceable rights or responsibilities” (FDA 2014).”

• cites Sarker et al. 2015Not checked yet

  Not checked in the free run.

• cites FDA 2014Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Because the literature conflicts, we cannot make definitive claims about our findings' validity.

“These conflicting findings in the literature prevent us from drawing conclusive claims about the validity of our findings.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Geddes and colleagues reported no substantial efficacy difference between SSRIs and TCAs.

“Geddes et al. found no major differences in the efficacy of SSRIs and TCAs”

• cites Geddes et al.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Other studies reported that one antidepressant class outperformed others.

“whereas other studies found one kind to perform better than others (Cipriani et al. 2018).”

• cites Cipriani et al. 2018Not checked yet

  Not checked in the free run.

May combine two claims.

Psychiatric medications are still chosen by trial and error or by their side effect profiles.

“in psychiatry, medications are still prescribed by trial-and-error, or based on side effect profiles of these medications (Trivedi et al. 2006).”

• cites Trivedi et al. 2006Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Imipramine and Clomipramine, both TCAs, are rarely prescribed now due to severe toxicity that can be fatal in overdose.

“While TCAs (Imipramine, Clomipramine) are not often prescribed today because of serious toxicity issues that may be fatal in overdose (Kerr, McGuffie, and Wilkie 2001)”

• cites Kerr, McGuffie, and Wilkie 2001Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The FDA aims to rely more on real-world evidence alongside pre-market clinical trial data.

“The FDA seeks to advance its approach to regulate and rely more on real-world evidence in addition to pre-market clinical studies data.”

• cites fda.govNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Our findings provide a tool for generating real-world evidence that regulators such as the FDA could use.

“Our results offer an important tool in generating “real-world evidence” for incorporation into technologies that can be used by regulatory bodies like the FDA.”

May combine two claims.

Unlike clinical trials, social media analysis cannot closely monitor treatment, so the number of measurements per person varies widely.

“Last, social media analysis does not allow us to closely monitor the treatment, unlike a clinical trial, which results in high variance in the number of measurements that each individual contributes.”

May combine two claims.

Using first-person singular and auxiliary verbs are known traits often seen in depressed people.

“the known characteristics of these two attribute usage, typically prevalent in depressed individuals (De Choudhury et al. 2013).”

• cites De Choudhury et al. 2013Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Some studies found placebos or non-drug care did better than certain antidepressants.

“Other studies found placebos or non-pharmacological care to have outperformed certain antidepressants (Szegedi et al. 2005).”

• cites Szegedi et al. 2005Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Medication self-report bias can cause treatment leakage: some controls may use drugs without mentioning it on Twitter.

“Additionally, self-reporting bias about medications can lead to treatment leakage, where some control individuals may be taking medications, but not mentioning it on Twitter.”

May combine two claims.

Propensity score matching accounted for many confounders, but other hidden factors-such as duration, history, dosage, compliance, and additional treatments-could still affect the outcomes.

“While we considered many confounders in our propensity score matching approach, there are other latent factors that could impact the effects considered here; e.g., duration, history, dosage, and compliance of using self-reported medications; additional medications or adjuvant treatments one might be using.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Although we match users on several traits, we do not preselect people likely to benefit from a drug.

“Second, even though we match users based on several characteristics, we do not pre-qualify individuals as potential beneficiaries of a medication.”

Classified as probably not a substantial claim.

The paper presents a new technological approach that could serve the FDA's changing needs by detecting medication uses and effects from self-reports on social media.

“This paper offers a novel technological approach that may meet the evolving needs of the FDA, by being able to identify the uses and effects of various medications as self-reported by people on social media.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Our work offers a new dynamic perspective on mental health, capturing real-time variation and incorporating dynamic systems, network, and instability theories.

“our work represents a novel dynamic viewpoint onto mental health— limitations notwithstanding, it captures the real-time variation and accounts for dynamic systems theory, network theory, and instability mechanisms (Nelson et al. 2017).”

• cites Nelson et al. 2017Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

We caution against making direct clinical inferences.

“Without dampening the clinical potentials, we caution against making direct clinical inferences.”

Probably not a substantial claim: may not pass S2 (checkable).

This new perspective contrasts with the traditional static view of psychiatric medication effects.

“Such a new window onto the field clearly contrasts the traditional static viewpoint on the effects of psychiatric medications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A clinician coauthor identified the most salient patterns from the regression coefficients using clinical literature and experience.

“We summarize a few distinct patterns below, noted by our clinician coauthor to be most salient, based on the clinical literature and experience:”