Identifying medications underlying communication atypicalities in psychotic and affective disorders:A pharmacosurveillance study within the FDA Adverse Event Reporting System

Background: Disruptions in language and speech are considered promising markers of affective and psychotic disorders. However, little is known about the mechanisms and confounders underlying such communicative atypicalities. Medications might have a crucial, relatively unknown role both as potential confounders and relatedly offering an insight about the mechanisms at work. The integration of regulatory documents with pharmacovigilance techniques could provide a more comprehensive picture to account for in future investigations of communication-related biomarkers. Objectives: Our aim was to identify a list of drugs potentially associated with speech and language atypicalities within psychotic and affective disorders. Methods: To structure a search for potential drug-induced communicative atypicalities on the FDA Adverse Event Reporting System (FAERS, updated June 2021), we developed a query using the Medical Dictionary for Regulatory Activities (MedDRA). We performed a Bonferroni corrected disproportionality analysis (Reporting Odds Ratio) on three separate populations: psychotic, affective, and non-neuropsychiatric disorders, to account for the confounding role of different underlying conditions. Unexpected drug adverse event associations, which were not already reported in the SIDER database of labeled adverse drug reactions, were subjected to further robustness analyzes to account for expected biases. Results: We identified a list of 291 expected and 91 unexpected potential confounding medications. We corroborated known/suspected associations: e.g., corticosteroids-related dysphonia and immunosuppressant-related stuttering. We also identified novel signals: e.g., domperidone-associated aphasia or VEGFR inhibitors-related dysphonia. Conclusions: We provide a list of medications to account for in future studies of communication-related biomarkers in affective and psychotic disorders. The methodological tools here implemented for large scale pharmacosurveillance investigations will facilitate future investigations of communication-related biomarkers in other conditions and provide a case study in more rigorous procedures for digital phenotyping in general. Objectives: Our aim was to identify a list of drugs potentially associated with speech and language atypicalities within psychotic and affective disorders, to account for in future investigations of communicative markers and to provide tools for similar future endeavors. Methods: We identified terms from the Medical Dictionary for Regulatory Activities (MedDRA) related to speech and language adverse drug reactions and clustered them by partial semantic overlap to structure a search on the FDA Adverse Event Reporting System (FAERS, updated June 2021). A Bonferroni corrected disproportionality analysis was applied to three separate populations in the FAERS: psychotic, affective, and non-neuropsychiatric disorders, to account for the confounding role of different underlying conditions. Unexpected drug adverse event associations, which were not already reported in the SIDER database of labeled adverse drug reactions, were subjected to further robustness analyzes to account for expected biases. Results: We identified a list of 291 expected and 91 unexpected potential confounding medications. We also developed methodological tools for large-scale pharmacosurveillance investigations: a MedDRA query proposal for speech and language impairments, formalization of possible biases, and related analyzes to account for them. Conclusions: We provide a list of medications to account for in future studies of communicative behavioral biomarkers in affective and psychotic disorders. The developed methodological tools will facilitate future investigations of communicative biomarkers in other conditions and provide a case study in more rigorous procedures for digital phenotyping in general.


The confounding role of medications on communication-related biomarkers
Affective and psychotic disorders have long been associated with atypical communicative patterns -e.g., decreased emotional expression and flat prosody 1,2 .
This awareness is widely used during the assessment of the disorders, and is increasingly investigated through automated voice and content analysis [3][4][5][6][7] . The combination of new powerful forms of machine learning, pervasive smartphone data collection, and other sources of big data will allegedly identify historically elusive markers for affective and psychotic disorders and therefore enable more reliable diagnoses, continuous evaluation of symptoms, and perhaps even personalized treatment [8][9][10][11][12] . However, communication is a complex phenomenon and its relation to specific disorders is not straightforward, with many potential confounders and ethical considerations 6,13,14 .
Medications, which can be disproportionately associated with neuropsychiatric diagnoses and their co-morbidities, can affect not only mental health but also speech (more related to voice and prosody) and language patterns (more related to content).
For example, commonly used medications with anticholinergic effects (e.g., antihistamines and antidepressants) can cause reduced salivation flow (xerostomia) and sedation of the mouth, which could cause dysphonia and difficulty in articulation 15 .
As another example, high D2R occupancy antipsychotics are administered to patients with psychotic disorders, but have also been shown to associate with slower speech and increased pauses 16 . Therefore, it is often not clear whether the communicative atypicalities identified as behavioral biomarkers of affective and psychotic disorders could be at least partially confounded by medications. Unfortunately, more general investigations of the associations between communicative atypicalities and medications are still sparse, and no comprehensive overview is available (see Supplementary Material 1 -Section A for an overview of studies assessing the effect of medication on speech patterns in schizophrenia).
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint Therefore, the objective of the current study was to identify a list of drugs that could be associated with atypicalities in speech and language, which should be evaluated in the future as potential confounders in communication-related biomarkers of affective and psychotic disorders. We first (Section 1.2) introduce our two key sources of information: clinical-trial-based information (SIDER 17 ) and spontaneous reports (FAERS 18 ). We then (Section 1.3) briefly discuss four common causal mechanisms underlying observed associations between drugs and adverse events. We present how the potential biases thus highlighted can be accounted for in our analyses (Section 1.4) before detailing our materials and methods (Section 2). Finally, we report and discuss the resulting list of drugs associated with communicative atypicalities (Sections 3 and 4).

Information sources
As medications are tested in clinical trials, adverse drug reactions are evaluated, and if the drug is approved for market distribution (marketing authorization) these adverse reactions are reported in the insert of the medication package 19 .
However, as the drug is used outside of clinical trials (post-marketing phase) unexpected drug reactions are often detected. For example, an adverse drug reaction could arise in populations not investigated in clinical trials (e.g., older or younger cohorts, pregnant women, patients with additional comorbidities). In addition, multiple drugs are often administered together (polytherapy), and an adverse drug reaction could arise from their interaction. Such suspected adverse reactions to drugs can be spontaneously reported to the regulatory agencies by physicians, marketing authorization holders, and the general public. Pharmacosurveillance is the method that critically collects these spontaneous reports to provide a more comprehensive safety profile of medications 20 .
Clinical trials and pharmacosurveillance have complementary strengths.
Clinical trials have obvious advantages, primarily that, by carefully selecting homogeneous samples and randomly distributing them across interventions, they remove many possible confounders and provide a strong causal assessment.
However, the weaker evidence provided by pharmacosurveillance can cover a much broader variety of patients and drug uses. In addition, spontaneous reports that inform pharmaco-surveillance can contain adverse reactions that are commonly . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint underreported during clinical trials. For instance, rashes are easy to observe, and arrhythmias could be fatal, and therefore, both are relatively prominent in clinical trial reports 21,22 as compared with symptoms such as raspy vocal quality, or mispronunciations of speech sounds. However, communication impairments can be disabling from the patients' and their families' point of view, and therefore be more likely to be spontaneously reported, as has been shown for stuttering [23][24][25][26] .
Pharmacosurveillance has long acknowledged that spontaneous reports provide very noisy information riddled with well-known biases. For example, reports may be incomplete or duplicated, lack quality control of the information provided (e.g., patients do not have the right language and knowledge to accurately label their symptoms), contain potential biases, and may ignore external factors such as the novelty of a drug and how media coverage of adverse reactions affects the number of reports 19,27,28 . In other words, causal connections between drugs and adverse reactions should not be established based solely on pharmacosurveillance data.
Nevertheless, by taking these biases into account, pharmacosurveillance can generate hypotheses for further investigation by analytical studies (cohort and casecontrol studies). Finally, with large enough sample sizes, there are methods for approximately estimating the causal effect of drugs in observational studies by adjusting for these newly considered confounders through confounding-adjustment methods 29 such as propensity score matching.

Causal models underlying drug adverse event associations
Pharmacosurveillance approaches identify adverse events that are more frequently present in reports about a given drug than in reports not containing that drug. However, the observed association could be generated through different causal mechanisms, with four common ones represented in Figure 1. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint (in the center). The diagrams are direct acyclic graphs (DAGs), that is, graphs in which the nodes (ellipses) are the observable phenomena, and the arrows are the causal connections (which can only be acyclical, that is, go one direction and not form loops). DAG A represents the case in which the event is an actual Adverse Reaction caused by the administration of the drug of interest. DAG B represents a case of Reverse Causality, in which the drug is administered to treat the adverse event but is incorrectly reported. DAG C represents a case of Confounding by Indication, in which the underlying condition that justifies the use of the drug also more frequently present the adverse event. DAG D represents a case of Confounding by Concomitant, in which the adverse event is a reaction to a coadministered drug (administered for the same condition or a related comorbidity).
The first possible causal model is simply that the adverse event is indeed caused by the drug (an Adverse Reaction to it; DAG A). For example, administering anticholinergic drugs often results in reduced salivation flow (xerostomia) and sedation of the mouth, which can cause speech impairment 15 .
However, the association might also result from Reverse Causality (DAG B): the drug is taken because of the event (e.g., to treat it) 1 . For example, botulinum toxin is approved to treat spasmodic dysphonia, and antipsychotics are administered offlabel to reduce stuttering 30 . These drugs can be reported as associated with a speech impairment because, for example, the lack of specific fields for symptoms of the underlying condition or for comorbidities often generates ambiguity in the reported information. Furthermore, when therapy does not reduce symptoms, reports might 1 While the adverse event is causing the prescription of the drug, the drug itself could be affecting the symptom (e.g., diminishing it) and therefore a more nuanced causal model than this simplified DAG would have to include bidirectional causal arrow, or a temporal dimension to causation. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint incorrectly record the indication for use (pre-existing stuttering) as an adverse reaction (after drug administration, stuttering is still there).
A third common possibility is the so-called "common cause" or fork 31 . Here the underlying condition is causing both the prescription of the drug and the adverse event, without there being any direct causation between the latter two (Confounding by Indication; DAG C). For example, psychotic disorders can involve some degree of communication impairment (e.g., alogia, i.e., reduced and vague speech, or disorganized speech), as well as the administration of antipsychotics. Therefore, when assessing all reports on FAERS, one might find an association between communication impairments and antipsychotics simply due to their co-presence, even if there were no direct causal association. Another example of the "common cause" problem is seen with gastroesophageal reflux, for which proton pump inhibitors (PPI) are administered. Acid reflux can also affect the larynx and vocal cords, resulting in dysphonia 32 , which would then appear to be associated with PPI even in the absence of a direct causal link.
A fourth common possibility is that the adverse event is indeed an adverse reaction, but to a different concomitant drug also prescribed due to the underlying condition (Confounding by Concomitant; DAG D). For example, diuretics are usually administered in conjunction with angiotensin-converting enzyme inhibitors (ACEI), which are known to cause bradykinin-related cough and laryngeal irritation. Therefore, diuretics might appear to be associated with dysphonia, even if the latter were exclusively due to ACEIs.
Finally, the relationship between a drug and an event may also not be reducible to one DAG only. Botulinum toxin may indeed be used to treat spasmodic dysphonia (DAG B), but it was also subject to a warning by the FDA because the systemic spread of the toxin can lead to flaccid paralysis and related dysphonia (DAG A) 33 .

From causal models to statistical analyses
When pharmacosurveillance identifies an association between a drug and an adverse event, how do we discriminate between the possible causal mechanisms? It turns out that there is no replacement for clinical and scientific knowledge: evidence . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint from previous studies, clinical expertise, and informed mechanisms hypotheses. This knowledge must play a meta-statistical role in guiding the construction of statistical analyses. In other words, it is up to clinically and scientifically informed pharmacosurveillance analyses, not statistics alone, to identify plausible directions of causality and the necessary follow-up studies.
Specifically, reverse causality (DAG B in Figure 1 Figure 1). This is the case of psychotic and affective disorders being associated with communicative impairment (e.g., flat prosody for both types of disorders, and semantic incoherence for psychotic disorders). A solution to this bias is to explicitly include the common cause in the model ("blocking the backdoor path" 31 ), for instance, by analyzing the populations separately: in our case, this implied separately analyzing individuals with affective disorders, individuals with psychotic disorders, and individuals without any neurologic medication in order to test whether patients with, e.g., affective disorders on vs. off a specific drug display higher rates of the adverse event of interest. By looking at reports for individuals not assuming any neurologic medication, we exclude (and therefore can correct for) is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint While these techniques provide information about potential mechanisms, they do not guarantee accurate causal inference. Nevertheless, they contribute to the collective construction of more accurate knowledge on the relationship between drugs and communicative impairment by providing hypotheses to be explored and assessed in future investigations.

Overview of the analyses
The general pipeline of the analysis is represented in Figure 2, the details of which are explained in the following paragraphs.

Definition of search terms
We relied on two information sources: SIDER for clinical trial reports of adverse drug reactions and FAERS for spontaneous post-marketing reports. Both sources employ a standardized hierarchical lexicon to code for adverse events, the Medical . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. Dictionary for Regulatory Activities or MedDRA. In MedDRA, the highest organization level is the System Organ Class (SOC, e.g., nervous system vs. vascular disorders), followed by the High-Level Group Terms (HLGTs, e.g., neuromuscular vs. neurological disorders), followed by High-Level Terms (HLTs, e.g., muscle tone abnormal vs. motor neuron disease) and Preferred Terms (PTs, e.g., hypertonia vs. hypotonia). Both SIDER and FAERS code their adverse events as preferred terms.
The MedDRA lexicon has some limitations. First, it does not always include the most adequate preferred terms to report a given adverse event; therefore, some events are less likely to be reported or are reported relying on only partially relevant terms. At the same time, preferred terms are quite detailed, and reporters are not always aware of the differences between terms (e.g., dysarthria and dyslalia), leading to their indiscriminate use. Thus, good pharmacovigilance practices involve so-called Standardized MedDRA Queries (SMQs), which are validated search queries that aggregate many partially overlapping preferred terms to identify and retrieve cases of interest. In the absence of an SMQ for speech and language disorders, we selected the preferred terms concerning communicative impairment. To increase understanding of the results, we clustered strongly overlapping terms, which were then traced, when possible, to separate Speech and Language categories. The use of generic preferred terms (e.g., speech disorder) often implies low expertise in the reporter and could therefore indicate any speech or language impairment.
Consequently, we chose to consider them separately, not to contaminate the other analyses. This clustering was performed independently by six clinical and domain experts (pharmaco-surveillance experts, speech-language pathologists, psychologists, and experts on voice markers; see the list of co-authors), and disagreements were discussed until resolved. The aggregated list consists of nine main categories (sub-queries) and 16 isolated terms. It is represented in Figure 3 and it is available in the Supplementary Material 1 -Section B -Table S1.

The FDA Adverse Event Reporting System (FAERS)
The United States Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) collects worldwide spontaneous reports of suspected adverse drug reactions and offers the highest accessibility to the public for customized analyzes. Specifically, its raw quarterly data include demographic, therapeutic, and . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022.

Exposure of interest
In order to identify medications associated with communicative impairments and deal with possible "common cause" biases (DAG C in Figure 1), we separately investigated three clinical populations: patients with a) affective, b) psychotic conditions, and c) without any neurologic medications (i.e., likely without any neuropsychiatric conditions). To identify patients with psychotic and affective conditions we selected all reports that recorded any preferred term (PT, for example, 'schizophrenia') belonging to the high-level group terms (HLGT) for psychotic disorders ("schizophrenia and other psychotic disorders") and affective disorders ("manic and bipolar mood disorders and disturbances" and "depressed mood disorders and disturbances"). To identify patients without neuropsychiatric conditions, we selected all reports that did not include any neurologic drug (according to the ATC) nor any psychotic or affective preferred term. The results of the selection procedure are displayed in the Supplementary Material 1 -Section C - Figure S1. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint potential biases not a priori acknowledged. For example, if we find that older people are more represented in cases than noncases, this may point to a potential bias related to a higher frequency of speech disorders in the elderly.

Disproportionality analyzes for drug-event association detection
Disproportionality analysis (the analysis of a reliably more frequent reporting of an adverse event in presence of a drug than in presence of any other drug, Figure 2 -Step 1) was performed following good pharmacovigilance practice guidelines 28 .
Using a contingency table 2x2, we calculated the Reporting Odds Ratio (ROR) whenever at least 10 cases of the event investigated (e.g., phonatory preferred terms) co-occurred with a drug (e.g., testosterone). The ROR was deemed significant when the lower limit of its 95% confidence interval was greater than 1. In other words, we report a potential association when the adverse event is more likely to be reported together with the drug of interest than with any other drug but the one analyzed.
We performed a disproportionality analysis evaluating associations between drugs (from the ATC 2022 classification, excluding mineral supplements and drugs included in the 'Various' class) and communication-related adverse events (subclusters of overlapping terms as identified in Figure 3). The analyses were run on all reports involving a) affective and b) psychotic disorders, and c) non-neuropsychiatric reports. To filter out likely spurious associations, results were subjected to Bonferroni correction.

The Side Effect Resource (SIDER)
The Side Effect Resource is a public database that grants free access to the information contained in the package inserts, that is, the official information on a drug and its uses, in particular its side effects, compiled and distributed by the drug manufacturers. Package inserts are text-mined, and the information retrieved is coded using the ATC classification for medications and the MedDRA classification for adverse events.
For each subquery of potential adverse reactions, we searched the specific preferred terms. We considered the identified medications as expected associations (Figure 2 -Step 2), which did not require further discussion of potential . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint biases and causal mechanisms. The associations found in FAERS but not present in the SIDER were considered unexpected and further assessed for potential causal confounding (Figure 2 -Step 3).

Aggregating Results
The expected adverse drug reactions from the SIDER and robust unexpected associations from the FAERS were aggregated in nine lists, one per each main subquery (e.g., phonatory issues or incoherence), Figure 2  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint antipsychotics, antihistamines, and antidepressants; see Figure 4, and the public repository 35 for a collated heatmap at the level of single active ingredients).

MedDRA query for case retrieval
We defined a MedDRA query with nine main categories: three concerning speech (phonatory, prosody, motor control or execution), six concerning language (aphasia, high/low productivity speech and thought, stereotypies, incoherence, abnormal reasoning) (see Figure 3). Note that some of the selected preferred terms could not be included in the larger clusters: seven more idiosyncratic language terms and nine more generic terms, which are not included in the subsequent analyses.  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint

Populations of interest
We selected three populations of interest: 302,000 reports involving affective disorders, 11,631 psychotic disorders, and 7,703,183 non-neuropsychiatric disorders.
A detailed presentation of the number of cases (reports with communication-related adverse events) and noncases is presented in the Supplementary Material 1 -Section C Figures S1-S5 and Table S2.

Expected and Unexpected Solid Associations
In the three populations, we performed disproportionality analyses for each category (see Supplementary Material 2). We detected both expected and unexpected drug associations (see Figure 4) and performed robustness analyses on the latter ones.
The result was a list of 291 expected and 91 unexpected potential confounding medications. Emerging results were shown in Supplementary Material 1 -Section C and Table 1. No association was found for prosody or abnormal reasoning. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Overview
Given the increased focus on communication-related biomarkers of affective and psychotic disorders, we need a more careful overview of how medications could potentially act as confounders. We developed a rigorous pipeline combining evidence from drug package inserts with post-marketing pharmacosurveillance tools and relied on causal inference techniques to account for potential biases. We identified a list of known (given the package insert) and unexpected medications that could affect communication-related biomarkers.
In the following subsections, we discuss how to interpret and use these findings and methods in the broader context of digital phenotyping trying to identify markers of neuropsychiatric conditions. First, we discuss expected and unexpected potential adverse reactions as they relate to the specific context of communication-related biomarkers of psychotic and affective disorders. Second, we present the limitations and possibilities of our approach. Finally, we discuss possible realistic uses of the list in future research. In other cases, we found unexpected associations with drugs from already known classes (integrated findings, that is, drugs from the same class were already known to associate with the adverse reaction). For instance, we found evidence for haloperidol being related to adverse reactions concerning motor control or execution (e.g., impacting articulatory precision) that are not reported in its package insert.

Known and unexpected adverse reactions
However, these reactions are reported in other antipsychotic package inserts and may . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint be a more general class effect. This is also the case for the unexpected signal concerning clonazepam (an antiepileptic, also used to treat anxiety) being associated with aphasia and for antineoplastic agents (mainly VEGFR-inhibitors) with dysphonia.
Other associations are even more unexpected. Medications used to treat cancer, such as plant alkaloids, cytotoxic antibiotics, protein kinase inhibitors, and monoclonal antibodies, emerge as potential causes of aphasia, which are not reported in the SIDER database. Crucially, since there is at least some evidence of increased cancer risk in schizophrenia 39 , we could expect a more common use of these drugs in patients with schizophrenia than in controls. Therefore, the adverse reaction could influence how well a predictive model could detect psychotic disorders from speech or language patterns, at least in complex machine learning models. Nevertheless, these drugs have never been mentioned -to our knowledge -in previous studies of vocal markers as possible confounders.

Formalized query
In the attempt to retrieve cases of interest in the FAERS, we found an oftenambiguous lexicon covering communicative impairments. Our effort to explicitly formalize a MedDRA query is a necessary step to focus the attention and to create a common framework for pharmacosurveillance research on these impairments.
The current query presents some limitations. For instance, one might more closely investigate how physicians describe and report these impairments. For example, common terms used by physicians to report dysphonia are acute laryngitis, nonspecific dysphonia, benign vocal fold lesions, and chronic laryngitis 40 , and for retrieving antipsychotic-related dysarthria cases one may search also for extrapyramidal syndrome and laryngospasm. More work is needed to cover these labels and validate the results of searches that integrate them.
Perhaps more crucially, we observed a high proportion of communicationrelated FAERS cases submitted by the general public. This suggests that communicative adverse events might be at the same time underplayed by medical practitioners, and of crucial importance to patients, caregivers, and families. In fact, we observed that patients with communicative impairment tend to specify the resulting disability more frequently in their reports than patients with other adverse events but . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; the same underlying condition. This suggests that future involvement of patient perspectives and the development of a better MedDRA lexicon and corresponding definitions for speech and language preferred terms might be an important step.
One could also question whether FAERS' and SIDER information is sufficiently sensitive to the kind of properties analyzed in the search for communication-related biomarkers. For example, minor articulatory impairments (e.g., increased jitter, that is, low-level irregularities in voice pitch) -useful in predictive, machine learning algorithms 41 -could not be perceived, or at least not perceived as enough of an issue, by patients and clinicians to be reported and precisely labeled.
Nevertheless, the construction of an initial query enables initial explorations of medication-based confounders and facilitates proposals, thus representing an important step in the development of a useful Standardized MedDRA Query (SMQ).

Causal inference
Although still uncommon in pharmacosurveillance, formalized causal inference, and the use of DAGs, in particular, are promising endeavor 31,42 . These tools provide a standardized framework for the formalization, visualization, and communication of confounding. These tools also provide structured and more reproducible procedures to account for at least some of the biases when designing analyses 31 .
We have built four relatively simple DAGs of the mechanisms underlying observed drug event observations. Thus, we have tried to identify the most problematic biases for our questions and accordingly adjusted our analyses and interpretation. However, it is important to note that many biases could not be fixed and that the characteristics of the reporting (often incomplete and unverified) complicate attempts at causal inference. For example, proton pump inhibitors are always used for gastroesophageal reflux, which may be responsible for laryngitis and dysphonia, and therefore the causal direction cannot be easily identified. Further, our broad focus did not permit us to delve into the richness of pharmacosurveillance data (e.g., information on concomitants, therapy regimen, co-occurring events) and to map more complex scenarios (e.g., variables affecting at the same time the use of the drug, the incidence of the adverse event, and the reporting of it). For example, botulinum toxin has been referred to as a potential cause and treatment for dysphonia, and more research is needed to disentangle these possible scenarios. In addition, biases, such as notoriety . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint bias, and masking bias, adjustment for the Weber effect 27 , are beyond the purpose of this study but should be considered when investigating specific drugs more closely.

Integrating additional sources
The main objective of spontaneous reporting systems is to collect useful data to identify unexpected associations between a drug and an adverse event in a timely and cost-effective manner. This identification enables early intervention and therefore limits the costs of drug-related harm. To effectively target currently not known, it is extremely important to integrate already acquired knowledge, which may come from the literature or from regulatory sources, primarily package inserts (FDA) and Summaries of product characteristics (EMA).
Databases that store this information in an easily accessible way are a promising tool for large-scale analyses because reading each individual package insert is time consuming. The SIDER uses a natural language processing algorithm to extract the information from regulatory sources and has not been updated since 2016 17 , therefore, it plausibly contains errors and outdated information.
A worldwide database in which data for each marketed drug is compiled and regularly updated by the marketing authorization holder and stored in an accessible way would enrich both regulatory activities and pharmacosurveillance analyses. In the meantime, the use of the SIDER or similar databases may help in large-scale analyses to reduce the risk of classifying already known reactions as unexpected signals.
We acknowledge that we cannot be sure whether some of the unexpected associations have already appeared as notes in clinical trials (but not reported in the package insert) or in subsequent scientific literature. Future work could attempt to integrate these additional sources of information. However, independent of the novelty, our list aggregates large amounts of otherwise dispersed information in an easier to consult format.
Future work could integrate additional sources of information (e.g., scientific literature) and provide weights to different sources according to the degree of evidence available (e.g., via Bayesian analysis).

Large-scale analyses
Traditional pharmacosurveillance studies focus on at most a handful of drugs and/or adverse events 43,44 . Thus, they can provide a fine-grained analysis of potential . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint Large-scale analyses require a broader overview, which cannot match the same level of detail and discussion. The strategies we implemented to simultaneously assess large sets of adverse events and drugs may help design future large-scale analyses. These strategies range from correction for multiple testing and automatic integration with regulatory databases, to an attempt to formalize possible underlying causal mechanisms and the use of a priori expected biases to implement robustness analyses. Large-scale analyses, however, provide only an initial perspective and must be complemented with more detailed studies of specific associations and their confounds.

How should this list be used?
We advocate for the list of drug confounders ( Figure 4 and Table 1)-whether as a cause of a speech or language atypicality or as a proxy of an underlying susceptibility -to be used in future studies of communication-related behavioral biomarkers by either including the presence of a medication as a covariate, removing participants who take medication, or interpreting results and study limitations as a function of which medications were taken. We know that, as observed by multiple reviews, most studies of such markers involve small sample sizes 2,46,47 . Such studies would be at a loss trying to adjust for such a large number of medications and would lack reliable evidence related to all but the most commonly used ones. Although a single study can still check the list for the most likely confounders (e.g., much higher use of drug x in the target population than in the controls), the real potential lies in the cumulative aggregation of this information across studies. The key is to report medications used by participants in individual studies, which would allow future megaanalyses 48 (aggregating datasets across studies preserving individual-level data) to directly assess the impact of a large variety of relevant medications.
Accounting for confounders is also important in machine learning studies. Current reviews and perspectives on the study of communication-related behavioral biomarkers advocate the collection of larger and more diverse samples and the use of state-of-the-art machine learning techniques, such as deep learning 6,7,14 . In these contexts, the algorithms can detect even the presence of weak confounding if it improves prediction. In other words, many machine learning models are likely to . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint recognize small differences between groups they try to classify. If these differences are due to higher levels of medication being used and not due to the target disorder, the models may not generalize well to other samples of the disorder where the medication use is different, which is common when changing countries and sociodemographic settings. Accordingly, a deeper understanding of the confounders and mechanisms at work is a key component also for more data-driven machine learning approaches, for instance, to guide bias assessment or even to identify more rigorous pipelines (e.g., presenting medication-balanced validation sets).
Finally, this list may also help identify more general hypothesized mechanisms underlying adverse events beyond a specific drug. Pharmaco-surveillance can thus act not only as a guide for precautionary regulatory action, but also as a hypothesis generation tool for scientific research, which could lead to follow-up studies involving, e.g., electronic health records (to assess adverse events before and after drug administration), experimental setups, and clinical studies. For instance, a more thorough investigation of the association between domperidone and aphasia would be of particular interest, given the biological plausibility -i.e., its activity as a dopamine antagonist-and the existence of conditions that increase the blood-brain barrier permeability. This might lead to more generalizable predictions as regards confounding drugs and increased understanding of the communicative features of the disorders over time.

Applications of the methods to other neuropsychiatric conditions
In the current study, we have currently focused on affective and psychotic disorders since previous research explicitly called for better investigation of medication-related confounders in identifying communication markers for these populations 1,2,5 . However, with proper consideration, the list could be easily extended when assessing communication-related behavioral biomarkers for other conditions. Neurodevelopmental conditions (e.g., autistic spectrum disorder), and neurological diseases (e.g., Parkinson's disease) seem obvious follow-ups.

Conclusions
Motivated by the increasing interest in communication-related behavioral biomarkers of affective and psychotic disorders, we set out to investigate the potential role of medications in affecting communication-related markers of these disorders. We is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint extracted the drugs already expected to cause communicative impairment from the SIDER. This paved the way for a pharmaco-surveillance analysis of a larger set of communication-related adverse events and drugs, controlling for prominent biases.
We corroborated known/suspected associations: e.g., corticosteroids-related dysphonia and immunosuppressant-related stuttering. We also identified novel signals: e.g., domperidone-associated aphasia or VEGFR inhibitors-related dysphonia. The results are combined in a list of medications to be accounted for in future studies on communication and biobehavioral markers of affective and psychotic disorders.
The work showcases methodological innovations to facilitate large-scale pharmacosurveillance approaches and identifies current shortcomings, along with discussing potential causal and pathogenetic mechanisms. In particular, the existing lexicon to identify communicative adverse events is underarticulated, perhaps due to an underappreciation of the perspectives of patients. We advocate for future work on this.
Drugs that confound the effect between communication-related behavioral biomarkers and psychiatric disorders are abundant. There should be concern not only for confounding drugs and comorbidities, but also non-medical substances and habits (e.g., smoking, vocal use). Here, we provide a tool for learning about and potentially adjusting for the confounders to improve digital phenotyping research. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted September 6, 2022. ; https://doi.org/10.1101/2022.09.05.22279609 doi: medRxiv preprint