Article Text
Abstract
Introduction Language is usually considered the social vehicle of thought in intersubjective communications. However, the relationship between language and high-order cognition seems to evade this canonical and unidirectional description (ie, the notion of language as a simple means of thought communication). In recent years, clinical high at-risk mental state (CHARMS) criteria (evolved from the Ultra-High-Risk paradigm) and the introduction of the Clinical Staging system have been proposed to address the dynamicity of early psychopathology. At the same time, natural language processing (NLP) techniques have greatly evolved and have been successfully applied to investigate different neuropsychiatric conditions. The combination of at-risk mental state paradigm, clinical staging system and automated NLP methods, the latter applied on spoken language transcripts, could represent a useful and convenient approach to the problem of early psychopathological distress within a transdiagnostic risk paradigm.
Methods and analysis Help-seeking young people presenting psychological distress (CHARMS+/− and Clinical Stage 1a or 1b; target sample size for both groups n=90) will be assessed through several psychometric tools and multiple speech analyses during an observational period of 1-year, in the context of an Italian multicentric study. Subjects will be enrolled in different contexts: Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), Section of Psychiatry, University of Genoa—IRCCS Ospedale Policlinico San Martino, Genoa, Italy; Mental Health Department—territorial mental services (ASL 3—Genoa), Genoa, Italy; and Mental Health Department—territorial mental services (AUSL—Piacenza), Piacenza, Italy. The conversion rate to full-blown psychopathology (CS 2) will be evaluated over 2 years of clinical observation, to further confirm the predictive and discriminative value of CHARMS criteria and to verify the possibility of enriching them with several linguistic features, derived from a fine-grained automated linguistic analysis of speech.
Ethics and dissemination The methodology described in this study adheres to ethical principles as formulated in the Declaration of Helsinki and is compatible with International Conference on Harmonization (ICH)-good clinical practice. The research protocol was reviewed and approved by two different ethics committees (CER Liguria approval code: 591/2020—id.10993; Comitato Etico dell’Area Vasta Emilia Nord approval code: 2022/0071963). Participants will provide their written informed consent prior to study enrolment and parental consent will be needed in the case of participants aged less than 18 years old. Experimental results will be carefully shared through publication in peer-reviewed journals, to ensure proper data reproducibility.
Trial registration number DOI:10.17605/OSF.IO/BQZTN.
- Adult psychiatry
- Child & adolescent psychiatry
- Health informatics
- EPIDEMIOLOGY
- Protocols & guidelines
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
Strengths and limitations of this study
Using validated diagnostic criteria, the study aims to improve the characterisation of early psychopathology with the results of a fine-grained analysis of language, hoping to define proper linguistic biomarkers.
The selection of relevant linguistic features is performed through a data-driven approach, without predefined cut-offs correlated to pathological significance.
Spoken-language data can be highly variable, and a key challenge concerns the optimisation of such a phase of data acquisition to allow the extraction of relevant information during the subsequent phase of textual processing.
The retention of participants for the entire duration of the observational period is a further challenge.
The assessment of conversion to full-blown psychopathology during the second year of prolonged observation may represent a methodological issue.
Introduction
Language, thought and human beings
Language is usually considered the social vehicle of thought in the context of intersubjective communications. This canonical interpretation of the thought-language relationship implicitly entails the priority of the first term. Therefore, in common medical practice, the verbalisation of delirious contents during an acute psychotic episode is notoriously considered as the manifestation of an underlying thought disorder.1 On the other hand, interspecies cognitive studies revealed the existence of a fundamental gap between the expression of linguistic and non-linguistic contents.2–4 The relationship between language and high-order cognition seems to evade the canonical unidirectional description and the common idea of a thought primacy. In fact, it is at least plausible that language acquisition exhibits a critical role for human cognitive development5: a subsequent deficit in cognitive functions has been experimentally linked to a primary insufficient development of linguistic skills.6–8 Moreover, according to Tattersall’s theory,9 10 language should be regarded as the fact that permitted the transition across different cognitive phenotypes during the evolution of human species. In philosophy, many authors highlighted the deep constitutive character of language for humans.11 12
Language and phenomenal experience
As well known, consciousness definition is still an ongoing issue. Among other things, it could be simplistically represented as an active and fundamental background that synthetise experienced phenomena within a spatiotemporal schema. This said, the linguistic apparatus, analysable through advanced natural language processing (NLP) techniques, can be considered as something that shapes the product of this primary synthesis to further reduce the phenomenal complexity and to allow the emergence of a unique and well-defined subjective experience. Therefore, the conscious phenomenon offers itself as a synthetic elaboration of an unrefined experience, following a double logic (first transcendental and then formal-linguistic). In this context, language stops being exclusively a means of communication for predefined thoughts, thus becoming a refined tool of interaction with the experienced world, regardless of what it can be said about abstract thought function. In some recent interpretations this complex problem has been transposed within the theoretical framework of predictive brain13 14 to show that ‘verbal cues (even if self-generated) can act as highly flexible (and metabolically cheap) contexts (set of priors), to generate a predictive signal helping the system to process an input that is otherwise too weak or noisy”.15
Psychopathology and language: floating on fluid psychopathological substrates
Most relevant psychopathological conditions appear to originate during early life stages.16 17 At the same time, the boundaries between previously distinct psychopathological disorders seem to weaken.18–20 The inadequacy of classical nosography has become progressively more evident, especially when it is applied to the dynamicity of early psychopathological phenomena.21 22 A classification based on strictly defined diagnostic categories appears inadequate when considering the complex phenomenon of comorbidities23 24 and symptomatological overlaps, frequently expressed during early stages of mental illness.25–29 To address the complex world of early psychopathology the so-called Ultra-High-Risk paradigm has been proposed over the past decades,30 originally developed to individuate schizophrenia prodromes.31 More recently, the concept of at-risk mental state has been expanded to detect conditions of ‘trans-diagnostic risk’.32–35 Coherently, Hartmann and colleagues36 proposed a new methodology based on the application of clinical high at-risk mental state (CHARMS) criteria and enriched with the introduction of a clinical staging system37 (table 1). Specifically, the detection of the CHARMS criteria allows to identify adolescents and young adults (aged up to 25 years old) expressing a psychopathological condition of transdiagnostic risk, which corresponds to stage 1b in the context of the clinical staging system (CS 1b). Coherently with the transdiagnostic approach, the risk is referred to a generic ‘exit syndrome’, that is, a first episode of full-blown psychopathology, defined through the overcoming of some psychometric thresholds, as well as through the verification of DSM-5 (Diagnostic and Statistical Manual of Mental Disorders - Fifth Edition) criteria for a specific disorder. Within the wide group of CHARMS+ / CS 1b subjects, some subcategories of ‘attenuated syndromes’ have been proposed by original authors (see table 2): the psychosis trait vulnerability group; the bipolar trait vulnerability group; the attenuated psychotic symptoms group; the attenuated (hypo)manic symptoms group; the moderate (attenuated) depression group; the attenuated borderline personality group; the brief limited intermittent psychotic symptom group. An ongoing observational study has been structured starting from this latter proposal, aimed to verify the predictive power of some novel risk category defined through the application of CHARMS criteria.38 Preliminary results have recently been published.39 To conceive our proposal and to first verify data reproducibility, we initially acquired the methodology described by Hartmann and colleagues.36 However, we chose to further enrich the experimental apparatus to gather different information from a fine-grained analysis of the linguistic profile, derived from subjects’ speech. This analysis will be conducted on textual data, revised by researchers, after a first automated transcription of the audio records, directly acquired during experimental assessment. In fact, language, as previously reinterpreted, can represent a further pathoplastic/pathogenetic factor, favouring the pathological crystallisation of phenomenal data.
Language production disturbances in psychosis and schizophrenia have been investigated since some early works promoted by Harrow and Quinlan40 and Andreasen and Grove,41 the sudden development of innovative methods of automated linguistic analysis promoted further investigations (for a comprehensive review see Corcoran et al 2020,42 increasingly oriented towards the early stages of the disease. In fact, it seems that language alterations may soon represent valid and practical biomarkers to perform a multilayered assessment of psychotic risk and to offer more tailored interventions. The hope concerns the possibility of extending this approach within the abovementioned transdiagnostic risk paradigm.
NLP techniques and their application in neuropsychiatric conditions
As reported by Voleti and colleagues,43 it is possible to identify different levels of complexity in linguistic analysis. For each level, several features have been proposed as potentially relevant in association with different neuropsychiatric conditions.
Lexical level
This level allows to extract features that account for the diversity and richness of lexicon used in a text. At this level, the following metrics are usually computed: (a) the type/token ratio (TTR), a standard index of lexical variety; (b) the Moving Average type-token ratio,44 considered as an ‘advanced’ TTR as it calculates lexical variety of a sample using a moving window that estimates TTRs for each successive window of fixed length. Moving Average TTR, Brunét Index, Honoré’s Statistic, part-of-speech (POS) tagging, aim to quantify lexical diversity and density. These parameters have been mainly studied either for risk stratification or for the diagnosis of morbid conditions of purely neurological relevance.45–49 These variables are quick and easy to assess. However, this simplicity is reflected in a reduced capacity of providing relevant information.
Morpho-syntactic level
This level allows to extract information from the POS-tagging step of linguistic annotation. In particular, the following variables are usually computed (a) the percentage distribution of morpho-syntactic categories (both functional and lexical); (b) the ‘lexical density’ index (ie, the proportion between functional words over the total number of words in a text).
Syntactic level
At this level whole propositions are examined to analyse the way in which words are organised in sentences and sentences in speeches. Considering the work of Mota and colleagues,50 the researchers extrapolated objective parameters of language measurement, useful for quantifying the alterations characteristically found in specific morbid states. The set of verifiable syntactic features covers a wide range of properties which can be further grouped. For instance, features related to the parse tree structure (eg, maximum parse tree depth, average length of dependency links), to the use of specific syntactic relations (eg, use of coordination and subordination) and to canonicity effects (eg, relative order of subject and object with respect to the verb).
Semantic analysis
The analysis of linguistic expressions in relation to the meaning they acquire in speech. Among related NLP methods, one of the first developed is the Latent Semantic Analysis (LSA),51 today carried out through the application of specific Machine Learning algorithms, exploiting artificial neural networks word2vec or GloVe.52 53 These tools can probabilistically define, starting from the analysis of large textual corpora, the semantic content of individual words and develop a specific vocabulary. More recently, further algorithms have been developed that can operate similarly at the level of entire propositions (eg, sent2vec, InferSent, Universal Sentence Encoder - USE). One of the first studies carried out in this area aimed to measure, by applying LSA, the semantic coherence of the language of patients suffering from Formal Thought Disorder (FTD) of different severity.54 Furthermore, through these methods the importance of semantic and pragmatic alterations in schizophrenia was confirmed.55 In recent years, the characteristics of language potentially predictive for the onset of psychosis in clinically defined high-risk subjects (clinical high-risk, CHR) were also investigated.56 57 Similarly, Rezaii and colleagues58 defined a so-called digital phenotype useful for quantifying the risk of psychosis onset in CHR subjects. According to Morgan and colleagues,59 different NLP measures may provide complementary information, being potentially associable to distinct aspects of mental disorders.
Aims and objectives
According to our theoretical speculations and inspired by previous works proposed by Hartmann and colleagues,36 39 we designed a multicentre observational study with the following objectives:
Primary objective
To investigate the alterations of multiple spoken language variables in subjects at pluripotential risk (CHARMS+ and Clinical Stage 1b) of developing a range of full-blown psychopathological disorders by estimating the associations between those variables and the conversion to full-blown psychopathological disorder in reference to a second group (CHARMS− and CS 1a), internally defined after the exclusion of the presence of a full-blown disorder (CS 2), as well as of a transdiagnostic risk condition (CHARMS+ and CS 1b).
To prospectively confirm the predictive and discriminant validity of the CHARMS criteria in a sample of CHARMS+ and CS 1b subjects and in a second group (CHARMS− and CS 1a).
Secondary objectives
To develop a prediction model for the probability of conversion to full-blown disease (Clinical Stage 2), using CHARMS criteria, CHARMS subgroups, linguistic features and a data-driven subset of language markers.
To evaluate the predictive and discriminative capabilities of such model.
Crucially, the conversion to full-blown conditions (ie, Clinical stage 2) is referred to a set of ‘exit-syndromes’, in line with the original methodology (ie, psychotic disorder, bipolar disorder, depressive disorder and borderline personality disorder).
Methods and analysis
Participants and setting
We designed a longitudinal follow-up study with an observational period of 2 years. The expected sample size is n=180: 90 subjects who meet CHARMS criteria (Group A: CHARMS+; CS 1b) and 90 controls (Group B: CHARMS−; CS 1a). Potential participants are help-seeking people aged 14–25 who are referred to local mental health service. Subjects will be enrolled in different contexts:
Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), Section of Psychiatry, University of Genoa—IRCCS Ospedale Policlinico San Martino, Genoa, Italy;
Mental Health Department—territorial mental services (ASL 3 – Genoa), Genoa, Italy;
Mental Health Department—territorial mental services (AUSL – Piacenza), Piacenza, Italy.
Group A (CHARMS+ and CS 1b): attenuated syndrome with moderate/subthreshold symptoms.
Inclusion criteria are: (1) the verified presence of a pluripotential ‘at-risk’ mental state—CHARMS+ and Clinical Stage 1b attenuated syndrome; (2) age between 14 and 25 Years; (3) native speakers (Italian) with a good understanding of spoken and written Italian language; (4) the ability to give informed consent by participants themselves and/or by parental authority holders in the case of a minor.
Exclusion criteria are: (1) a documented history of intellectual disability or a diagnosis of autism spectrum disorder; (2) current or previous diagnosed full-blown mental disorder—Clinical Stage ≥2; (3) the presence of relevant neurological disorders (including brain trauma, epilepsy, stroke, cerebral palsy); (4) a mental condition directly and exclusively induced by substances or by organic causes (in accordance with DSM-V criteria); (5) the identification through psychic examination of delusional persecutory thoughts, such that the experimental audio recording may be a source of discomfort for the enrolled subject.
Group B (CHARMS− and CS 1a): mild or non-specific symptoms.
Inclusion criteria are: (1) the exclusion of a pluripotential ‘at-risk’ mental state—CHARMS+ and Clinical Stage 1b attenuated syndrome; (2) age between 14 and 25 Years; (3) native speakers (Italian) with a good understanding of spoken and written Italian language; (4) the ability to give informed consent by participants themselves and/or by parental authority holders in case of a minor.
Exclusion criteria are: (1) a documented history of intellectual disability or a diagnosis of autism spectrum disorder; (2) current or previous diagnosed full-blown mental disorder—Clinical Stage≥2; (3) the presence of relevant neurological disorders (including brain trauma, epilepsy, stroke, cerebral palsy); (4) a mental condition directly and exclusively induced by substances or by organic causes (in accordance with DSM-V criteria).
Procedure and data acquisition
Baseline interview A—psychometric measures
For each enrolled subject, the necessary informed consent is acquired (for participants aged <18 years, parental consent will also be obtained). The baseline examination will be conducted by preselected and pretrained research team members and will involve a general neuropsychiatric evaluation and a structured recording of medical/family history. The following psychometric scales will be administered to finally verify inclusion/exclusion criteria and subjects eligibility:
**Comprehensive Assessment of At-Risk Mental States38—semi-structured interview, seven different psychopathological domains rated according to a global rating score (0–6), a frequency score (0–6) and a substance use score (0–2). Italian validation.60
**Structured Clinical Interview for DSM-5 (SCID-5)61 and SCID-5 Personality Disorders (SCID-5_PD)62—semi-structured interviews for establishing clinical diagnoses (gold standard) based on DSM-5. Regarding the SCID-5_PD section, only the modules concerning borderline personality disorder and schizotypal personality disorder are considered for the purposes of this study.
**Social and Occupational Functioning Scale (SOFAS)63—observer-rated (0–100) scale assessing the social and occupational functioning.
Global Functioning Scale: Social and Role (GFS/GFR)64 65 (clinician rated)—both deriving from Global Assessment of Functioning, GFS assesses (from 1—extreme dysfunction to 10—superior functioning) quantity and quality of social relationship, while GFR assesses (same scoring system) subject’s performance in different contexts (school, work or home).
Quick Inventory of Depressive Symptomatology-Clinicians rated (QIDS-C)66—clinician-rated 16-items questionnaire that assess the severity of depressive symptoms during the previous week.
**Young Mania Rating Scale67—11 clinician-rated items that assess (gold-standard)68 severity of manic symptomology over the previous 48 hours. Italian version.69
Depression Anxiety Stress Scale 21 (DASS-21)70—self-report scale (short version of the 42-item DASS) to assess three domains (seven items for each domain) of negative affectivity referred to the past weeks. When the scale is administered in children and adolescents, only one (general) score is defined.71
Bipolar Spectrum Diagnostic Scale (BSDS)72—self-rating narrative-based scale which assesses the entire bipolar spectrum, including subthreshold states of bipolar illness.73
**Personality inventory for DSM-5, brief version74—self-rating screening75 76 tools (25 items) for assessing in adult and adolescents five maladaptive personality trait-dimensions, described according to the alternative model of personality disorder.
Davos Assessment of Cognitive Biases Scale (DACOBS)77 78—self-report scale with 42 items to assess the presence of possible cognitive biases, cognitive limitations and avoidance behaviours.
Munich Cronotype Questionnaire (MCTQ)79—a self-report tool to assess information on sleep referred to work and work-free days and to quantitatively obtain a chronotype related to sleep intervals.
**Insomnia Severity Index80 81—self-rating (seven items) instrument to assess night-time and daytime symptoms of insomnia in adults and adolescents.82 Italian version.83
** Italian version available.
As indicated, some of these psychometric tools (ie, QIDS-C, SOFAS, GFS/GFR, DASS-21, BSDS, DACOBS and MCTQ) have not been officially validated in Italian, thus a preliminary internal translation was realised to reproduce as much as possible a certain methodology36 and to eventually perform an internal validation of the abovementioned psychometric tools.
If a participant will exceed the threshold for a full-threshold disorder (Clinical Stage ≥2), then he/she will be excluded from the study and committed to the mental health service. On the contrary, if the subject will meet the CHARMS criteria (CHARMS+ and CS 1b) or falls below the CHARMS threshold (CHARMS− and CS 1a), then he/she will be included in the research programme (unless additional exclusion criteria are met).
Through the acquisition of psychometric variables and the application of CHARMS criteria it is also possible to verify for each subject in Group A the subgroup of risk, as proposed by Hartmann et al36 (table 2).
Baseline interview B—speech recording
The baseline assessment includes a second part to be carried out after a few days (T0-b), according to a shared agenda and by the same research team members.
At T0-b, subjects of both groups will be first evaluated through the Montreal Cognitive Assessment scale.84 85 Besides, the spoken language of enrolled subjects will be audio-recorded. Participants will describe four sequences of vignettes, picturing four logically linked events. In each sequence human individuals engaged in simple actions within contexts of daily life are represented. Two sequences were specifically created to be affectively neutral; a third one should be more emotionally salient; finally, a fourth sequence should express a less intuitive logic of transition between the single depicted moments. Then, they will be asked to answer four predefined questions, each related to a particular detail of each picture in a sequence. The free speech of each participant will also be recorded, eventually elicited with some questions, formulated according to narrative interview’s recommendations (ie, phenomenological inquiry paradigm86
The recording sessions will last 30–45 min. Data will be acquired using the same recording device and using a free software (Auphonic), with the following settings:
Format: CAF/WAV (PCM).
Sample rate: 44 khz or 48 khz.
Channel: MONO.
Depth: 16 bits.
Time series
The first year of observation includes three phases of data acquisition for each participant.
During each phase, all relevant information about participants will be registered/updated. Linguistic data will be recorded at Tn-b, following the same methodology of acquisition described for T0-b. Data acquired at each time point are summed in table 3.
Crucially, at each different phase of the first year, subjects’ conversion to full-blown disorder (Clinical Stage 2) will be verified. Each participant will be also periodically evaluated by mental health specialists (not directly involved in the study), who will eventually provide him/her with any appropriate therapeutic intervention. During the second year of observation subjects will be specifically assessed for conversion to Clinical Stage 2 any time a significative worsening of psychological status will be reported from the abovementioned standard periodical evaluation performed by external mental health teams.
Linguistic data processing and elaboration
At each phase of data acquisition, audio reports will be automated transcribed verbatim under the supervision of dedicated researchers. A first database of anonymous raw transcripts will be produced. In such a form, transcripts will be shared with our partner institution (Computational Linguistic Institute ‘Antonio Zampolli’, National Research Council (CNR), Pisa, Italy) to perform textual data processing. Starting from the transcripts, linguistic analyses will be performed along different levels. Raw text and (morpho-)syntactic analysis will be automatically carried out by means of Profiling-UD,87 a multilingual web-based tool that provides a comprehensive assessment of language use. The tool performs a two-stage process: linguistic annotation (carried out by UDPipe)88 according to Universal Dependencies (UD) formalism89 and linguistic profiling. The annotated texts will be used as input to the further step, performed by the linguistic profiling component that defines the rules to extract and quantify the formal properties. The final output of the process is a vector-like representation that can comprises more than 120 linguistic features: (1) shallow features, for example, average length and counts of words and sentences, (2) morpho-syntactic features, for example, POS tagging distributions and inflectional properties of verbs or (3) more complex features obtained from syntactic parsing of the sentences, such as the use of subordination. The set of features from Profiling-UD has been derived from the literature on linguistic complexity, language acquisition and neurolinguistics, and have been successfully applied in a wide range of tasks and scenarios: from the automatic tracking of developmental patterns in child language acquisition90 91 and the evolution of written language competence in school learners,92 93 to the prediction of behavioural and cognitive impairments based on the detection of relevant linguistic markers from clinical tests.46 94
Furthermore, semantic representations of each transcript will be computed for both single sentence and the whole session level. To this end, we will rely on state-of-the-art neural network architecture, for example, Transformers models,95 which have shown massive improvements in NLP. Particularly, Natural Language Understanding models based on this technology, such as those of the BERT family96 have defined new states of the art in many tasks (eg, GLUE collection of benchmark tasks). The main advantage of these recent models over previous methodologies is that the embedding of a word is not fixed but computed for every occurrence based on its lexical contour; they can also applied in pathological contexts97. We plan to exploit a pretrained BERT model for the Italian language that has been trained on a huge corpus of more than 13 billions of words, that is, ‘bert-base-italian-cased’ to encode semantic information of words and sentences. Following Corcoran et al,42 we plan to analyse the coherence in the flow of subject speech by computing the semantic closeness of contiguous sentences (ie, the cosine distance between the embedding representations of the sentences).
Duration of the study
The recruitment kick-off is scheduled for July 2022; the database lock will be carried out according to the achievement of a sufficient sample size. The minimum expected duration of the observational period for a non-dropped out participant is 2 years. Preliminary data referred to each participant will be analysed at the end of recruitment phase. After this first passage data could possibly undergo a correction process due to the potential delayed conversion to full-blown disease (Clinical Stage 2) during the second year of observation.
Patient and public involvement
Research questions and outcomes were defined to address the complexity of early psychopathology and to better correspond to help-seeking young people needs, frequently expressed in real-world mental health settings. The richness of patients’ symptoms descriptions and their expressive urgency guided the development of the study design, prompting us to focus our attention on the characteristics of spoken language. At the end of the period of data acquisition, the results of the experimental investigations could help inform patients’ primary advisers, potentially optimising the care offer. Furthermore, during the individual assessment, the exceeding of predefined psychopathological thresholds (conversion to full-blown disease—Clinical Stage 2) will be communicated to the patient’s advisers to immediately adopt appropriate therapeutic measures.
Statistical analyses
Estimated sample size and statistical power
As reported by Hartmann and colleagues,36 literature'-based expectations of the 1-year transition rate in the CHARMS+ and CHARMS− groups are, respectively, 20% and 3%. To detect such a 6.7-fold increase as significant with 90% power, 5% significance level and 20% drop-out rate, a total of 180 subjects are required. Hence, we defined an expected sample size of n=180 (n=90 group A and n=90 group B).
Power calculations were performed by simulation with R software V.4.0.2. All the supporting material is available at a publicly repository.
Statistical analyses post data acquisition
Primary analysis will determine whether the rate of patient’s conversion to full-blown disease (Clinical Stage 2) in CHARMS+ patient’s group differs from the rate in CHARMS− group. Pearson’s χ2 test will be performed on the 2×2 contingency table of patient’s group (CHARMS+ or CHARMS−) and the occurrence of the Stage 2 conversion over a fixed follow-up time (which is one or 2 years of observation in preliminary and final analyses, respectively).
In the presence of heterogeneity in drop-out rates between the two patient’s groups the main analysis will be performed both in the complete case data set and under the assumption of the conversion rate to 1 (most conservative approach) for all drop-out patients.
In the presence of significative imbalances of patient’s characteristics between the two patient’s groups a multivariable logistical regression analysis will be used to adjust for potential confounders. Both raw and adjusted analysis will be reported with OR, 95% CIs and p values. A sensitivity analysis considering the conversion event as a time dependent outcome will be preplanned to make full usage of all available follow-up periods and make each day of observation contribute to the final conversion rate estimation.
In the case of difference between the rate of Stage 2 conversion in CHARMS+ and CHARMS− patients three classification analyses will be performed to detect: (1) the clinical alterations, (2) the CHARMS subgroups and (3) the spoken language features most associated to the Stage 2 conversion regardless the CHARMS group. Within these three analyses variable selection procedures might be applied, such as: bidirectional stepwise, ridge, lasso or elastic net to identify the best combination of predictors to identify Stage 2 conversion. Receiver Operating Characteristic curve, the area below such curve and metrics derived from the 2×2 prediction-observation confusion matrix (such as sensitivity and specificity, positive and negative predictive values) will be used to estimate model’s predictive and discriminative capabilities. Internal validation procedures such as k-fold cross validation may also be required to improve model generalisability. If needed, external information source such as real-world prevalence rates may be used to contextualise model performance.
Due to the high number of tested, there is a high risk of labelling some false, spurious, associations as significative. Therefore, we define a three-step strategy aimed at mitigating this risk. First, we prospectively list a set of characteristics (N=41, online supplemental material) that will be tested for interaction with CHARMS group in Stage 2 conversion rate. Second, the p values to test these prespecified characteristic associations to the risk of Stage 2 conversion will be presented using Bonferroni correction for multiple comparisons based on the number of variables in the list (ie, 41, regardless of any data collection issues that may emerge during the study execution). Third, due to the data driven nature of the audio processing approach, other unplanned analyses on characteristics yet to be defined are expected to be performed; such analyses will be explicitly labelled as exploratory, and the reader will be acknowledged in the result presentation (and trough this protocol) to carefully look at the findings merely as ‘hypothesis generating’.
Supplemental material
All analyses will be performed with R software (or equivalent statistical software) and uploaded in a public repository to guarantee the transparency and replicability of any finding.
Ethics and dissemination
The methodology described in this study adheres to ethical principles as formulated in the Declaration of Helsinki and is compatible with International Conference on Harmonization (ICH)-good clinical practice. The research protocol was reviewed and approved by two different Ethics Committees (CER Liguria approval code: 591/2020 – id.10993; Comitato Etico dell’Area Vasta Emilia Nord approval code: 2022/0071963). Participants will provide their written informed consent prior to study enrolment and parental consent will be needed in the case of participants aged less than 18 years old. Experimental results will be carefully shared through publication in peer-reviewed journals, to ensure proper data reproducibility.
Ethics statements
Patient consent for publication
Acknowledgments
This work was developed within the framework of the DINOGMI Department of Excellence of MIUR 2018-2022 (Law 232/2016).
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
Twitter @lisa_marzano
Collaborators The LNG-PSY Study Investigators: Alessandra Costanza (Department of Psychiatry, Faculty of Medicine, University of Geneva (UNIGE), Geneva, Switzerland). Francesca Sibilla, Pietro Calcagno (Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), Section of Psychiatry, University of Genoa, Genoa, Italy. IRCCS Ospedale Policlinico San Martino, Genoa, Italy). Sara Patti, Gabriella Molino (Department of Mental Health and Pathological Addictions, Genoa Local Health Authority, Genoa, Italy). Andrea Escelsior (Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), Section of Psychiatry, University of Genoa, Genoa, Italy. IRCCS Ospedale Policlinico San Martino, Genoa, Italy). Alice Trabucco (Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), Section of Psychiatry, University of Genoa, Genoa, Italy). Lisa Marzano (Department of Psychology, School of Science and Technology, Middlesex University, London, UK). Dominique Brunato, Andrea Amelio Ravelli (Italian Natural Language Processing Lab, Institute of Computational Linguistics 'Antonio Zampolli' (ILC-CNR), Pisa, Italy). Marco Cappucciati (Department of Mental Health and Pathological Addictions, Piacenza Local Health Authority, Piacenza, Italy. Early Psychosis: Interventions and Clinical-detection (EPIC) lab, Department of Psychosis Studies, Institute of Psychiatry, Psychology & Neuroscience, King's College London, London, UK). Roberta Fiocchi, Gisella Guerzoni, Davide Maravita, Fabio Macchetti, Elisa Mori, Chiara Anna Paglia, Federica Roscigno, Antonio Saginario (Department of Mental Health and Pathological Addictions, Piacenza Local Health Authority, Piacenza, Italy).
Contributors The authors have all contributed to the manuscript equally. LMag and AAm conceptualised and designed the study. LMag, EMon, IT, AT and DMart wrote the first draft of the protocol. OB, SC, MI and GL carried out a first methodological integration to optimise the experimental design. GM, SP, PC, FS, MC, RF, GG, DMara, FM, EMor, CAP, FR and AS further corrected the design to adapt it to different experimental settings, included in the multicentric context. LC conceptualised the planning of statistical analyses. FD, DB and AAR from Italian Natural Language Processing Lab provided a detailed project for linguistic data extraction and analysis. MA, GS, LMag, LMar, AC, AAg and AE carefully revised and approved the final version of the manuscript. Furthermore, a special thanks goes to the patient’s primary advisers for their contribution in reporting enrollable subjects.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.