Schizophrenia has long been associated with disturbances of language and thought. Clinicians describe formal thought disorder, a pattern of disorganized or incoherent speech that can range from tangential rambling to outright derailment. Decades of research have linked this phenomenon to the cognitive deficits that many patients experience, including impairments in memory, attention, and executive function. What has been missing, the authors argue, is a systematic account of which specific, automatically extractable linguistic features track cognitive status in schizophrenia. Traditional clinical assessment of thought disorder is subjective, time-consuming, and dependent on the judgment of trained raters. Computational methods promise something different: objective, reproducible measurements that could one day support screening and monitoring in routine care.

To pursue that goal, the team assembled speech samples from 88 patients with schizophrenia and 69 healthy controls. Among the patients, 37 were classified as cognitively impaired and 51 as cognitively preserved, a distinction based on their cognitive test performance. This split is the heart of the study’s design. Rather than simply asking whether the speech of patients differs from that of healthy people, the researchers asked whether the speech of cognitively impaired patients differs from that of cognitively preserved patients, a far more demanding comparison that speaks directly to the question of whether language can serve as a window onto cognition in the illness.

The technical machinery behind the analysis was a set of natural language processing pipelines adapted specifically for Korean. Korean poses distinctive challenges for computational linguistics. It is an agglutinative language in which particles and endings attach to stems, word boundaries are less transparent than in English, and the order of elements within a sentence is relatively free. Standard NLP tools built for English cannot simply be repurposed. The researchers therefore adapted their pipelines to handle Korean morphology and syntax, and from the transcribed speech they extracted features across four broad categories: lexical, syntactic, semantic, and sentiment.

Each of these categories captures a different dimension of language production. Lexical features measure the richness and diversity of vocabulary, including how many different words a speaker uses relative to the total number of words produced. Syntactic features quantify the structural complexity of sentences, such as the depth of embedding and the length and elaboration of clauses. Semantic features assess coherence, the degree to which successive utterances stay on topic and connect meaningfully to one another, which is precisely the quality that deteriorates in formal thought disorder. Sentiment features capture the emotional polarity of the language, tracking whether speech is expressed in positive or negative terms. Together, these measures form a multidimensional profile of each speaker’s verbal output.

The results were striking. Patients with cognitive impairment showed marked reductions across all four domains compared with both cognitively preserved patients and healthy controls. Their vocabulary was poorer, their sentences structurally simpler, their discourse less semantically coherent, and their emotional expressiveness flatter. The cognitively preserved patients, by contrast, fell between the two extremes, showing some linguistic disturbances relative to controls but considerably less pronounced than those seen in the impaired group. This graded pattern is important because it suggests that linguistic measures are not merely detecting the presence of schizophrenia but are sensitive to the degree of cognitive compromise within the disorder, which is exactly what a useful cognitive marker would do.

The researchers also took care to control for a potential confound that plagues speech studies in psychiatry: medication. Antipsychotic drugs can produce motor side effects, and the severity of such extrapyramidal symptoms could plausibly influence speech production independently of cognition. By using the Extrapyramidal Symptom Rating Scale to account for these effects, the team strengthened the argument that the linguistic differences they observed reflect cognitive and linguistic processes rather than the pharmacological treatment itself. Several of the linguistic features also correlated with the severity of patients’ symptoms and with their performance on language tasks, providing convergent evidence that the computational measures were capturing clinically meaningful variation.

The machine learning component of the study was framed as an exploratory analysis, and its results carry an instructive lesson about the state of the field. The researchers trained classifiers to distinguish the patient groups from one another and from controls, and they observed a discrepancy between the mean performance estimated across the outer folds of cross-validation and the performance achieved on a held-out test dataset. In practical terms, the models did not generalize as well as the cross-validation estimates suggested. This kind of gap is a well-known hazard in studies with modest sample sizes, where models can inadvertently capitalize on idiosyncrasies of the training data. The authors’ transparency about this discrepancy is a model of scientific candor, and it underscores a broader point: computational psychiatry is still in the phase where careful validation matters more than headline accuracy figures.

Why does a Korean-language study matter so much? Most of the existing literature on linguistic markers in psychosis has been conducted in English, and it has been an open question whether the same features, reduced lexical diversity, simplified syntax, degraded coherence, would appear in a language with a very different structure. The findings support the cross-linguistic generalizability of these markers, suggesting that the linguistic footprint of cognitive impairment in schizophrenia is not an artifact of English grammar or English NLP tools. If the same computational signatures emerge in typologically distant languages, the case grows stronger that they reflect something fundamental about how disordered cognition shapes language, rather than something incidental to a particular language community.

The implications reach toward clinical practice. Cognitive impairment in schizophrenia is a powerful predictor of long-term functional outcomes, yet it is often not formally assessed, and there are no approved treatments that reliably reverse it. A speech-based screening tool, requiring nothing more than a few minutes of recorded conversation and automated analysis, could offer a low-cost, repeatable way to identify patients at risk, track cognitive change over time, and measure response to interventions in trials. The present study does not yet deliver such a tool, as the generalization gap in the machine learning results makes clear, but it establishes the essential groundwork: the features exist, they are measurable in Korean, they track cognitive status, and they can be extracted without subjective judgment. As speech corpora grow and models are validated across larger and more diverse cohorts, the conversation itself may become one of psychiatry’s most informative diagnostic instruments.

Subject of Research: Linguistic markers of cognitive impairment in schizophrenia identified through machine learning analysis of Korean speech

Article Title: Linguistic markers in schizophrenia patients with cognitive deficit: a machine learning study using Korean speech

Article References: Choi, S.-H., Li, L., Seo, J. H., Le, T.-H., Setiani, A., Odkhuu, S., Kim, W. S., Nazir, S., Piao, Y., & Chung, Y. C. (2026). Linguistic markers in schizophrenia patients with cognitive deficit: a machine learning study using Korean speech. Schizophrenia. https://doi.org/10.1038/s41537-026-00807-1

Image Credits: AI Generated

DOI: 10.1038/s41537-026-00807-1

Keywords: schizophrenia, cognitive impairment, natural language processing, machine learning, speech analysis, formal thought disorder, Korean language, linguistic markers, semantic coherence, biomarkers, psychiatry, cross-linguistic validity

A team led by Cecilia Sanjuan-Ortiz of King’s College London, together with colleagues spanning Spain and the United Kingdom, conducted a systematic review following a pre-registered protocol and PRISMA reporting standards. Their search swept across five scientific databases to identify longitudinal studies that followed individuals at clinical high risk and reported which baseline characteristics were associated with remission from the high-risk state. After independent screening and data extraction by four reviewers, the team settled on fifty-one studies encompassing 10,073 participants from fifteen countries. The participants were young, averaging 19.1 years at study level with a range from 13.8 to 28.4 years, and just under half were female. Follow-up periods stretched from as short as eight months to as long as six years, giving researchers a meaningful window into how these young people fared over time.

The headline statistic is striking: roughly one-third of individuals at clinical high risk achieve remission from their CHR-P state at follow-up. That figure reframes the entire clinical conversation. For decades, research in this field has fixated on transition to psychosis, treating the onset of a diagnosable disorder as the outcome that matters most. But the new synthesis, published in the journal Schizophrenia, argues that remission deserves equal billing. A young person whose attenuated symptoms fade and who returns to their previous level of functioning has, in a very real sense, dodged the clinical bullet, and understanding what makes that outcome possible could transform early intervention from a blunt instrument into a precision tool.

Methodologically, the review carries considerable weight. The authors assessed risk of bias using a modified Newcastle-Ottawa Scale, and the results were reassuring: the mean score across studies was 7.7, with 92.1 percent of the included studies rated as high quality. Because the studies differed so substantially in their designs, measures and follow-up windows, the team could not pool results statistically; instead they performed a narrative synthesis, weighing how consistently each candidate predictor appeared across independent cohorts. That consistency, rather than any single effect size, became the currency of the review. A factor that showed up again and again in different countries, different clinics and different decades carries more evidentiary weight than one reported once, however dramatically.

What emerged was a clear hierarchy of predictors. The most robustly replicated factor was baseline functioning: ten separate studies found that individuals who were doing better in their daily lives, at school, at work and in relationships, when they first presented were more likely to remit. Close behind came symptom severity. Seven studies replicated the finding that lower levels of attenuated positive symptoms at baseline predicted remission, while three studies pointed to the same relationship for negative symptoms, the flattening of emotion, motivation and speech that often proves more disabling than hallucinations or delusions. Two studies independently identified lower self-disturbance, the unsettling sense that one’s thoughts, body or sense of self are slipping out of one’s own ownership, as a predictor of recovery.

Cognition emerged as another pillar. Four studies replicated the association between higher baseline verbal learning and later remission, and two studies each supported verbal memory, working memory and visual memory as predictors. The pattern makes intuitive sense: cognitive reserves buffer the impact of emerging symptoms and support the problem-solving, planning and social navigation that recovery demands. But the review did not stop at psychology. Three studies found that a preserved mismatch negativity amplitude, an electrophysiological brain response that fires when the brain detects an unexpected deviation in a repetitive sound stream, predicted remission. The mismatch negativity has long been studied as a candidate biomarker in psychosis research, and its appearance here, in the context of recovery rather than deterioration, suggests it may index the integrity of the very neural prediction machinery that psychosis disrupts.

Perhaps the most futuristic finding concerns language. Two studies replicated the association between speech fluency and connectedness and remission, joining a growing literature in which computational analysis of what patients say, how coherent their narratives are, how semantically rich their word choices, how smoothly their sentences flow, provides a window into underlying neurocognitive function. The idea that a short recorded speech sample could help clinicians stratify risk is no longer science fiction; it is an active frontier, and this review gives it a foothold in the CHR-P field specifically. Alongside these replicated findings, the authors flagged emerging but less consistent evidence for peripheral neurochemical, genetic and psychosocial or environmental factors, domains that remain underexplored and underpowered in the existing literature.

The review is candid about its limits. Substantial heterogeneity across the fifty-one studies, in how remission was defined, how predictors were measured and how long participants were followed, prevented any meta-analytic pooling and limits direct comparability of results. The authors also note that potentially protective factors associated with favourable outcomes remain understudied relative to risk factors, a bias that mirrors the field’s historical preoccupation with deterioration. Most studies examined predictors of remission from the high-risk state itself, but the relationships between remission, transition to psychosis and long-term functional recovery are not always straightforward, and the review stops short of claiming that any single marker can be used clinically today.

Still, the practical message is clear and, in its way, radical. Remission in CHR-P is consistently associated with lower symptom severity at baseline, preserved functioning and better cognitive performance, with self-disturbance, linguistic features and electrophysiological markers emerging as promising supplementary signals. No single domain, demographic, clinical, cognitive, neurophysiological or linguistic, tells the whole story. The authors argue that translation into clinical practice will require integrating information across domains rather than relying on single-domain indicators, effectively calling for multivariable prediction models that weigh a young person’s functioning, symptom profile, cognitive test results and possibly their speech patterns together. For clinicians running early intervention services, that means the most informative assessment may be the broadest one. For the young people and families sitting in those assessment rooms, the review offers something rarer than a risk calculator: evidence that for a substantial minority, and possibly with the right support, a much larger share, the story does not end in psychosis at all. It ends in recovery, and science is finally learning how to see it coming.

Subject of Research: Predictors of remission in individuals at clinical high risk for psychosis

Article Title: Predictors of remission in individuals at clinical high-risk for psychosis: a systematic review

Article References: Sanjuan-Ortiz, C., Laherran-Cantera, N., Aymerich, C., Dully, J., Ahmed, L., Salazar de Pablo, G., & Cella, M. (2026). Predictors of remission in individuals at clinical high-risk for psychosis: a systematic review. Schizophrenia. https://doi.org/10.1038/s41537-026-00809-z

Image Credits: AI Generated

DOI: 10.1038/s41537-026-00809-z

Keywords: clinical high risk for psychosis, CHR-P, remission, psychosis, systematic review, attenuated psychotic symptoms, cognition, verbal learning, mismatch negativity, self-disturbance, speech analysis, early intervention

Isolated REM sleep behavior disorder, or iRBD, is a sleep condition in which people physically act out their dreams, sometimes shouting, punching, or kicking during the night. The reason it attracts so much attention from neuroscientists is that a large proportion of people with iRBD eventually develop Parkinson’s disease or a related disorder, often years or even decades after the sleep problem first appears. That makes people with iRBD an ideal population in which to study the prodromal phase of Parkinson’s, the long silent stretch before the classic motor symptoms of tremor, stiffness, and slowness of movement become obvious enough for a clinical diagnosis. If researchers can detect subtle motor decline during this prodromal window, they could identify candidates for future preventive treatments and measure whether those treatments actually slow the disease.

The study enrolled twenty-six healthy controls, twenty-two patients with iRBD, and twenty-three patients with early-stage Parkinson’s disease. Every fourteen days, for twenty-four months, each participant completed a set of tasks on a smartphone covering four distinct motor domains: speech, psychomotor performance, hand tremor, and a U-turn task. The speech tasks included sustained phonation, reading aloud, and rapid repetition of syllables, a test known as oral diadochokinesis. The psychomotor component involved tapping exercises, while the tremor task asked participants to hold the phone in a way that allowed its motion sensors to record any shaking of the hand. The U-turn task captured information about turning speed, an aspect of movement that is known to be affected early in parkinsonism. All of this was done remotely, without clinic visits, which is precisely what makes the approach attractive for large-scale trials.

The results, reported by Vojtěch Illner and colleagues including corresponding author Jan Rusz, reveal a striking asymmetry between the two patient groups. In the iRBD group, disease progression showed up in two places: a measure called monopitch derived from reading aloud, which reflects a flattening of pitch variation in the voice, and a slowing of turn speed in the U-turn task. Both of these changes reached statistical significance at the p < 0.05 level over the two-year follow-up. In the early Parkinson’s disease group, the clearest longitudinal signal was a shortening of the time to the first voice break during sustained phonation, again significant at p < 0.05. In other words, the voice and turning tasks carried genuine information about disease trajectory, while other measures behaved differently.

That difference matters because many of the smartphone measures proved excellent at telling groups apart at a single point in time but failed to track change over the two years. Several speech, psychomotor, and U-turn measures showed adequate cross-sectional sensitivity, meaning they could reliably distinguish people with iRBD or early Parkinson’s disease from healthy controls when everyone was compared at the same moment. Cross-sectional discrimination and longitudinal sensitivity are not the same thing, however. A measure can separate a patient from a control on any given day yet remain stable within an individual, making it useless for monitoring progression. The study’s two-year design allowed the team to separate these two properties, something shorter studies often cannot do.

Perhaps the most instructive finding of the study concerns the learning effect. When people repeat the same task every two weeks, they improve at it, and that improvement can mask or mimic true disease change. The researchers found a significant learning effect during the first year of monitoring in measures derived from oral diadochokinesis, the rapid syllable repetition task, as well as from tapping and hand tremor. In practical terms, participants got faster or steadier simply through repetition, and only after roughly a year did this practice effect plateau. For anyone designing a future remote monitoring trial, this is a crucial piece of guidance: measures derived from these tasks need either a stabilization period, careful modeling of the learning curve, or exclusion from progression analyses, otherwise the data will reflect practice rather than biology.

The technical achievement behind these results should not be understated. Modern smartphones contain high-precision accelerometers, gyroscopes, and microphones that can capture movement and voice at sampling rates more than sufficient for detecting the small changes that characterize early neurodegeneration. Speech analysis can quantify pitch variability, timing, and stability of phonation with millisecond and hertz-level resolution. Motion sensors can decompose a turning maneuver into speed and smoothness. By administering these tasks every fourteen days, the study generated a dense longitudinal record for each participant, far richer than what annual clinic visits could provide. This density is what allowed the researchers to observe gradual drifts in voice and turning measures that would otherwise be lost in the noise of day-to-day variation.

For the field of digital biomarkers in Parkinson’s disease, the study functions as a roadmap rather than a final answer. The authors describe their work as guidance for future remote smartphone trials that integrate diverse motor domains to monitor prodromal and early parkinsonism. The message is twofold. First, speech measures, particularly those capturing pitch characteristics during reading and voice stability during phonation, together with turning speed, appear to carry genuine longitudinal signal in the populations most at risk. Second, the tasks that seem most intuitive, such as tapping speed and tremor measurement, are the ones most vulnerable to practice effects in the first year, and therefore require the most careful handling. Trials aimed at disease modification, where the endpoint is slowing of progression, will depend on measures that are both sensitive to change and robust to repetition.

The broader implications reach beyond Parkinson’s research. Remote smartphone monitoring could eventually allow neurologists to follow patients continuously at home, reducing the burden of travel and capturing the real-world variability of symptoms that clinic assessments miss. For people with iRBD, many of whom are keenly aware of their elevated risk, a regular two-minute phone battery could offer an objective, quantified view of their own trajectory, replacing anxiety with data. And for pharmaceutical companies developing therapies intended to be given before symptoms appear, validated digital endpoints collected at home could dramatically shrink and shorten the trials needed to prove that a drug works. The present study, with its modest sample sizes and two-year horizon, is an early step on that path, but a carefully designed one.

There remain caveats worth keeping in mind. The cohorts were relatively small, with twenty to twenty-six participants per group, and the findings will need replication in larger and more diverse populations. The study also monitored participants for two years, a period in which only a subset of people with iRBD can be expected to convert to overt Parkinson’s disease, so the full predictive power of the smartphone measures for identifying who will convert remains to be established. Still, the core conclusion stands on solid ground: a consumer device carried in a pocket, paired with well-chosen tasks and a smart analysis plan, can detect meaningful motor change in the earliest stages of Parkinson’s disease, provided researchers respect the learning effects that come with repetition. As remote monitoring matures, the smartphone may well become the stethoscope of neurology, and studies like this one are teaching clinicians exactly where to listen.

Subject of Research: Smartphone-based remote motor monitoring for detecting disease progression in isolated REM sleep behavior disorder and early Parkinson's disease

Article Title: Two-year smartphone motor monitoring in isolated REM sleep behavior disorder and Parkinson’s disease

Article References: Illner, V., Novotný, M., Kouba, T., Tykalová, T., Šimek, M., Šubert, M., Sovka, P., Švihlík, J., Růžička, E., Šonka, K., Dušek, P., & Rusz, J. (2026). Two-year smartphone motor monitoring in isolated REM sleep behavior disorder and Parkinson’s disease. npj Parkinson's Disease. https://doi.org/10.1038/s41531-026-01588-1

Image Credits: AI Generated

DOI: 10.1038/s41531-026-01588-1

Keywords: Parkinson's disease, isolated REM sleep behavior disorder, digital biomarkers, smartphone monitoring, speech analysis, motor progression, prodromal Parkinson's, remote assessment, hand tremor, U-turn task, learning effect, npj Parkinson's Disease

The study’s central innovation is a new interview format called the Multi-Topic Emotion Interview, or MTEI. Rather than asking patients to complete a resilience scale, the researchers guided them through conversations spanning multiple emotional topics, recording both audio and momentary self-assessments during the session. Between December 2024 and November 2025, the team collected multimodal data from 248 patients diagnosed with major depressive disorder and 101 healthy controls. To test whether their models would hold up beyond the original cohort, they also ran a four-week longitudinal follow-up with 106 participants, creating a temporal validation set that mimicked the real-world challenge of applying a model to new patients weeks after it was built. The study was retrospectively registered with the Chinese Clinical Trial Registry under identifier ChiCTR2500100966 on April 17, 2025, and approved by the ethics committee of Beijing Anding Hospital.

At the heart of the analysis sits an expert-guided large language model, a design choice that distinguishes this work from most speech-based mental health studies. Instead of letting a general-purpose language model freely interpret transcripts, the researchers steered it with expert knowledge to extract two families of features. The first is a Semantic Resilience Index, or SRI, a quantitative score capturing how resilient a patient’s language sounds in terms of coping strategies and adaptive thinking. The second is a three-dimensional emotional profile known in the literature as VAD, standing for valence, arousal, and dominance, which characterizes the emotional tone of what a patient says along dimensions of pleasantness, intensity, and sense of control. These language-derived features were then combined with Self-Assessment Manikin ratings, a validated pictorial instrument in which patients report their momentary emotional state during the interview itself.

To turn these features into a continuous resilience score, the team employed a Voting Regressor ensemble, a machine learning technique that combines the predictions of multiple regression models to produce a more stable estimate. The target variable was the Connor-Davidson Resilience Scale, or CD-RISC, the standard clinical questionnaire for measuring resilience. The best-performing multimodal fusion model, which integrated the semantic resilience index, the VAD emotional features, and the SAM state assessments, achieved an R-squared of 0.555 and a Concordance Correlation Coefficient of 0.705. In practical terms, the model explained more than half of the variance in questionnaire-measured resilience, and its predictions agreed closely with the actual scores patients reported. Those are respectable figures for a behavioral biomarker, a field where effect sizes are often far weaker.

What makes the result scientifically interesting is how well the model traveled through time. When evaluated on the four-week follow-up data from 106 participants, performance barely moved: R-squared of 0.557 and a Concordance Correlation Coefficient of 0.701. In machine learning for medicine, models frequently degrade when applied to data collected later, a phenomenon driven by drift in recording conditions, patient populations, or clinical workflows. The near-identical metrics here suggest that the speech-derived resilience signal is stable rather than an artifact of a particular recording session. For a digital biomarker intended to be used repeatedly across the course of treatment, temporal stability is arguably as important as raw accuracy, and this study provides unusually direct evidence for it.

The researchers then asked a harder question: was the model actually measuring resilience, or was it merely detecting how depressed someone was? Because depression severity and resilience are related, a model could score well on resilience prediction while really just tracking mood. To disentangle the two, the team performed residual analysis, statistically removing the contribution of depression severity and examining what the model still captured. The answer was that the model explained 9.6 percent of the variance in pure resilience residuals, meaning it genuinely decoded resilience-specific information that exists independently of current depressive symptoms. This is the study’s most consequential finding, because it demonstrates that a speech-based system can identify a patient’s coping capacity even when that patient is in the depths of a depressive episode, when self-report instruments are least reliable.

Interpretability, often the Achilles heel of large language model applications in medicine, received explicit attention through SHAP analysis, a technique that quantifies how much each feature contributes to a model’s predictions. The analysis revealed that the most critical predictors were three LLM-derived constructs: agency and control, reflecting the degree to which patients describe themselves as capable of influencing outcomes; cognitive nuance, capturing the sophistication and flexibility of their thinking about difficult situations; and contextual emotional dominance, a measure of perceived control within emotionally charged content. These are precisely the psychological mechanisms that resilience theory says should matter, which lends the model a degree of face validity that purely data-driven features often lack. A clinician looking at the model’s output would not receive an opaque number but a profile grounded in recognizable cognitive dimensions.

Equally notable is what failed. Traditional acoustic features, the pitch, energy, and rhythm measures that have dominated speech-based depression research for decades, did not provide independent incremental validity during late fusion, the stage at which different feature streams are combined. In other words, once the model had access to the deep semantic features extracted by the expert-guided language model, the raw sound of the voice added essentially nothing. This challenges a common assumption in the digital psychiatry field that paralinguistic vocal cues are the primary carriers of diagnostic information. The finding suggests that for resilience specifically, what patients say and how they frame their experiences may matter more than how their voices sound, a conclusion that could redirect research funding and design choices across the field.

The clinical implications are considerable. A scalable, objective resilience measure could change how treatment is personalized in psychiatry. Two patients with identical symptom scores may have very different recovery trajectories depending on their coping resources, and today those differences are largely invisible to standard assessments. A speech-based biomarker could be administered during routine visits, requiring only a short interview and no additional hardware beyond a microphone, and could track whether an intervention is actually strengthening resilience rather than just suppressing symptoms. The authors suggest that the identified predictors, particularly agency and control and cognitive nuance, point to actionable targets for personalized interventions, since these are constructs that psychotherapies such as cognitive behavioral therapy explicitly aim to modify.

Cautions remain before such a tool reaches the clinic. The cohort was recruited at a single Chinese psychiatric hospital, and cross-cultural validation will be essential, since both language use and the expression of coping differ across populations. The follow-up period of four weeks is short relative to the relapsing course of major depressive disorder, and the study measured resilience against a self-report gold standard, albeit one that the model demonstrably exceeded in objectivity. The expert-guided design also raises questions about how sensitive the system is to the specific prompting framework used. Still, the combination of a large multimodal cohort, temporal validation, residual analysis, and mechanistic interpretability makes this one of the more rigorous demonstrations to date that language, parsed intelligently, can carry clinically meaningful psychiatric signal. As large language models continue to mature, the quiet details of how patients talk about their lives may become one of psychiatry’s most informative vital signs.

Subject of Research: Speech-based digital biomarkers of psychological resilience in major depressive disorder using expert-guided large language models

Article Title: Interpretable speech biomarkers of psychological resilience in major depressive disorder via expert-guided large language models

Article References: Wang, S., Zhu, X., Li, N., Yang, Q., Liu, Y., Wang, J., Liu, P., Feng, L., & Wang, G. (2026). Interpretable speech biomarkers of psychological resilience in major depressive disorder via expert-guided large language models. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08584-y

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08584-y

Keywords: major depressive disorder, psychological resilience, digital biomarker, large language model, speech analysis, natural language processing, Connor-Davidson Resilience Scale, SHAP analysis, multimodal fusion, psychiatric assessment, semantic resilience index, machine learning

The study drew on the Island Study Linking Aging and Neurodegenerative Disease (ISLAND), a ten-year community-based initiative in Tasmania, Australia, that combines dementia education with large-scale data collection. From more than 1,300 potentially eligible participants, the researchers assembled a cross-sectional sample of 339 cognitively unimpaired adults with an average age of about 66 years, roughly 71 percent of them women. Of these, 238 reported no noticeable cognitive changes and served as asymptomatic controls, while 101 met criteria for subjective cognitive decline (SCD)—a self-perceived, persistent worsening of memory or mental function despite normal performance on cognitive tests. SCD matters because older adults with it progress to dementia at roughly twice the annual rate of those without it, and neuroimaging studies suggest it shares pathological features with mild cognitive impairment and Alzheimer’s dementia.

The participants completed the Tasmanian Test, a browser-based, self-administered battery of motor, cognitive, and speech tasks that they performed at home without supervision. For the speech component, they recorded four maximally paced DDK tasks in fixed order: rapid repetition of the monosyllables “pa,” “ta,” and “ka,” followed by the trisyllable sequence “pataka.” Each test lasted ten seconds, with up to two reattempts allowed. From the resulting audio recordings, the team extracted a rich set of features quantifying repetitions, speaking time, pauses, voice onset time—the interval between the release of a plosive consonant and the onset of vocal fold vibration—and variability measures. Crucially, they also computed derived “motor cost” measures by subtracting performance on one stimulus from another, capturing differences in articulatory control between bilabial, alveolar, and velar consonants, and between simple and complex sequences.

Why should rapid syllable repetition be sensitive to Alzheimer’s disease at all? The neurobiological rationale rests on the brain circuits that govern sub-second timing and sequential motor control. Functional MRI studies show that paced syllable repetition engages the putamen and cerebellum in a rate-dependent manner, and that rapid syllable-sequence production recruits the cerebellum in proportion to articulatory complexity. Cerebellar-basal ganglia circuits are central to internal timing and error correction, and Alzheimer’s-related alterations in cerebellar connectivity have been documented. Unlike connected speech tasks such as picture description, which confound motor control with language and memory demands, DDK tasks isolate the pure motor dimension of speech production—making them a potentially cleaner window onto preclinical brain changes.

The researchers tested whether these motor speech measures were associated with three established Alzheimer’s risk indicators: subjective cognitive decline, carriage of the apolipoprotein E ε4 allele, and plasma concentrations of phosphorylated tau 181 (p-tau181). Each carries weight. APOE ε4 is the strongest genetic risk factor for late-onset Alzheimer’s, conferring roughly three- to fourfold increased risk with one allele and eight- to fifteenfold with two. Plasma p-tau181 rises during preclinical disease, performs comparably to tau PET for classifying amyloid positivity in people with SCD, and predicts progression to Alzheimer’s dementia in cognitively unimpaired individuals. Blood samples had been collected at in-person clinics across Tasmania in 2021, with p-tau181 measured using an ultrasensitive single-molecule array platform and APOE genotype determined by PCR.

The statistical approach was deliberately cautious. Bivariate, nonparametric analyses with Benjamini-Hochberg correction for multiple comparisons found that no associations with SCD or APOE ε4 survived adjustment—only one measure, a “pa”-“ta” cost of repetition duration variability, remained significantly linked to plasma p-tau181. The team then turned to LASSO-penalized regression, a machine learning technique that shrinks coefficients to prevent overfitting in high-dimensional data. To stabilize estimates against the randomness of cross-validation fold assignment, they used a percentile-LASSO procedure, running each tenfold cross-validation process one hundred times and fitting the final model at the 95th percentile of the regularization parameter. Logistic models predicted SCD and APOE ε4 status, while a linear model predicted natural log-transformed p-tau181, with age, sex, education, anxiety, and depression available as covariates.

The results were intriguingly patterned. In the main analysis of 315 participants with complete data, motor speech measures predicted SCD with an area under the curve of 0.700, APOE ε4 carriage with an AUC of 0.715, and log-transformed p-tau181 with a mean absolute error of 0.323. For SCD, the simple models selected only depression as a predictor, but the more complex “maximal” models—which allowed interaction terms—consistently outperformed them and selected several motor speech features, always in interaction with depression. For APOE ε4, the simple models prevailed, with voice onset time features and variability measures prominent among the selected predictors, alongside age. For p-tau181, age was the only non-speech covariate chosen, and one derived measure in particular—voice onset time variability cost for “ka”-“pataka”—appeared in the superior model of every analysis, sometimes standing alone and sometimes interacting with age, education, and depression.

Depression emerged as a recurring complication. Because depression can affect subjective cognition, speech, and motor function alike, the researchers ran sensitivity analyses excluding participants with Hospital Anxiety and Depression Scale scores of eight or above, and separately incorporating hearing loss and expanded samples of participants with incomplete biomarker data. These analyses revealed potentially mitigating effects of depression and hints of bias from the limited availability of blood biomarker data. In the expanded sample, APOE ε4 prediction fell to barely better than chance in the full group, yet improved markedly in the low-depression subset, where more motor speech predictors were selected. The authors interpret this as evidence that depression may modify or obscure genuine relationships between motor speech control and Alzheimer’s risk factors, and that even subclinical depression may need explicit adjustment in future studies.

The study has clear limitations that the authors acknowledge candidly. It was cross-sectional and exploratory, so temporal direction cannot be established and longitudinal follow-up will be essential. The sample was predominantly White, highly educated women of Northern European ancestry, limiting generalizability, and participants were health-motivated volunteers in a dementia risk reduction program. Exclusion of cognitive impairment relied on self-reported diagnoses, and blood biomarker data were available only for those who attended clinics, introducing potential selection bias that the expanded-sample analyses could only partially address. The authors also note that the differencing process used to compute motor costs may reduce intra-individual variability, complicating interpretation of those derived measures.

Even with those caveats, the findings broaden the Alzheimer’s motor phenotype—which already includes slower gait, reduced stride length linked to p-tau181, and faster motor decline in APOE ε4 carriers—into the domain of speech motor control. If replicated with counter-balanced task order and longitudinal designs, motor speech measures could become scalable, non-invasive behavioral indicators that complement gait analysis and blood biomarkers in population screening. The vision is compelling: a browser-based test that anyone can take at home, extracting subtle signatures of articulatory timing and consistency that reflect the earliest biological whispers of Alzheimer’s disease. For now, the study stands as careful, early-stage evidence that the voice may know something about the brain that standard cognitive tests do not yet reveal.

Subject of Research: Associations between motor speech performance in rapid syllable repetition tasks and Alzheimer's disease risk markers in cognitively unimpaired older adults

Article Title: Motor speech markers of Alzheimer’s risk

Article References: Motor speech markers of Alzheimer’s risk. (n.d.). https://doi.org/10.1007/s11357-026-02560-w

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02560-w

Keywords: Alzheimer's disease, motor speech, diadochokinesis, subjective cognitive decline, APOE ε4, p-tau181, blood biomarkers, digital biomarkers, dementia screening, GeroScience, speech analysis, preclinical detection

The name of the model encodes its architecture. ALSA stands for Alternating Local-Sparse-Atrous attention, a six-layer scheme that alternates three different forms of self-attention within a Transformer encoder. This encoder sits on top of convolutional neural network layers that extract fine-grained local spectral features from audio. The design was motivated, the authors explain, by two persistent weaknesses in existing audio-based depression detection algorithms: inadequate feature extraction and weak sequence modeling. Voices carry depression cues at multiple time scales simultaneously, from momentary changes in spectral texture to long-range shifts in emotional prosody, and a model must capture both to perform reliably.

Before any classification happens, the system performs standardized audio preprocessing and then extracts two complementary families of features. The first is a Mel-spectrogram computed via a sixth-order complex Gaussian Continuous Wavelet Transform, abbreviated cgau6 CWT. Wavelet transforms are well suited to speech because they analyze signals at multiple resolutions, zooming in on rapid transients while still characterizing slower spectral evolution. The second branch extracts dual-band Chroma features, which summarize the pitch content of the signal across musical pitch classes. Chroma representations are sensitive to pitch abnormalities, and depressed speech is known to exhibit altered pitch dynamics, reduced variability, and flattened intonation.

Once extracted, the two feature branches are fused into joint representations that capture both spectral and pitch abnormalities of depressive speech in a single embedding. This multi-scale fusion is a central theme of the paper. Rather than betting on a single feature type, the model combines information that responds to different aspects of vocal pathology. The fused representation then passes through one-dimensional convolutional layers, which act as local feature detectors, picking up short-term fine-grained spectral patterns that would be diluted if the raw sequence were fed directly to a global model.

The heart of the architecture is the six-layer Transformer encoder equipped with the alternating attention scheme. Standard self-attention in a Transformer weighs every position in a sequence against every other position, which is powerful but computationally expensive and can overemphasize irrelevant global relationships. The ALSA design instead cycles through local attention, which concentrates on neighboring frames and preserves short-term speech details; sparse attention, which restricts connections to a subset of positions to reduce redundancy and cost; and atrous attention, which skips positions in a dilated pattern to widen the receptive field without adding parameters. By alternating these mechanisms across layers, the encoder simultaneously captures short-term local speech details and long-range emotional temporal dependencies, the two scales that earlier models tended to handle poorly.

After the encoder, adaptive pooling compresses the sequence into a fixed-length vector, and fully connected layers perform the final binary classification: depressed or non-depressed. The pipeline is end-to-end, meaning that the network learns how to weight the fused acoustic features rather than relying on hand-crafted rules about what depressed speech should sound like. This matters clinically because vocal manifestations of depression vary considerably between individuals, and learned representations can adapt to that variability in ways that fixed acoustic heuristics cannot.

The performance figures reported on the Chinese EATD-Corpus, a publicly available dataset of speech from depressed and healthy controls, are the study’s headline result. The model achieves a precision of 0.9531, a recall of 0.9524, and an F1-score of 0.9525. Precision measures how many of the flagged cases truly are depressed, recall measures how many actual cases are caught, and the F1-score balances the two. Values above 0.95 on all three metrics indicate that the system rarely raises false alarms while still catching the overwhelming majority of true cases, a combination that matters enormously in a screening context where both missed diagnoses and stigmatizing false positives carry costs.

Robustness is a second key claim. Real-world recordings, whether made on phones in noisy homes or in busy clinics, are never clean, and models that perform well only on laboratory audio are of limited practical value. The authors report that ALSA-CNN-Transformer shows strong robustness to noise, which supports its proposed use for remote and large-scale screening. Because the method requires nothing more than a voice recording, it is noninvasive, objective, and cheap to deploy at scale, in contrast to interviews that demand clinician time and questionnaires that depend on honest and insightful self-assessment.

The clinical and social implications extend beyond raw accuracy. The researchers argue that an efficient, objective, noninvasive tool of this kind can assist depression diagnosis, promote early detection, and help reduce the stigma that still surrounds mental illness. A voice-based screen could be embedded in telehealth platforms, community health kiosks, or routine phone check-ins, flagging individuals who need fuller clinical evaluation before a crisis develops. Early intervention is one of the strongest predictors of better outcomes in depression, and scalable screening is precisely what current diagnostic infrastructure lacks.

The study also situates itself within a fast-moving research landscape. Recent work on audio-based depression detection has explored bidirectional LSTM networks with multi-head attention, graph neural networks applied to audio signals, convolutional autoencoders, and multimodal fusion models that combine audio with text and visual cues. Systematic reviews of speech-based depression detection have found that deep learning methods generally outperform traditional approaches, but also highlighted the need for better feature extraction and temporal modeling, exactly the gaps this new architecture targets. The Yanshan University team, led by Ailing Tan and corresponding author Yong Zhao, was supported by the Hebei Natural Science Foundation and the National Natural Science Foundation of China, and the underlying EATD-Corpus data are publicly available, with further data available from the corresponding author upon reasonable request. As voice-based AI health tools move from papers into practice, studies like this one suggest that the human voice may become one of the most accessible diagnostic signals in medicine, carrying in a few seconds of speech what questionnaires take pages to uncover.

Subject of Research: Audio-based depression detection using a hybrid CNN-Transformer deep learning model with multi-scale feature fusion

Article Title: ALSA-CNN-transformer: audio depression detection using hybrid attention and multi-scale feature fusion

Article References: Tan, A., Zhao, R., Ma, R., Wei, J., & Zhao, Y. (2026). ALSA-CNN-transformer: audio depression detection using hybrid attention and multi-scale feature fusion. Biomedical Engineering Letters. https://doi.org/10.1007/s13534-026-00617-5

Image Credits: AI Generated

DOI: 10.1007/s13534-026-00617-5

Keywords: depression detection, speech analysis, deep learning, CNN-Transformer, attention mechanism, wavelet transform, Mel-spectrogram, feature fusion, mental health screening, biomedical engineering, machine learning, EATD-Corpus

The toolkit, developed by Allegra Benzini, Alyssa M. Lanzi, Cinzia Palmirotta, Simona Aresta, Petronilla Battista, and Gaia C. Santi of the Istituti Clinici Scientifici Maugeri IRCCS and the University of Delaware, extends the CHAT transcription conventions and the CLAN analysis software developed within the CHILDES/TalkBank framework to Italian-speaking adults with neurological disorders. CHAT—Codes for the Human Analysis of Transcripts—and its companion program CLAN have long been the backbone of quantitative language analysis, implemented in 28 languages and applied to a database of more than 44 million spoken words. Yet the system was originally designed around English child language, and its conventions presuppose English grammar, English phonology, and English clinical presentation. The Italian adaptation, presented as a downloadable operational manual, systematically reshapes those conventions to fit the structure of Italian and the realities of aphasic speech.

The methodological heart of the work is a transparent decision tree applied to every symbol in the CHAT manual. Each of the 59 transcription symbols was evaluated symbol by symbol against explicit inclusion and exclusion criteria: symbols had to be relevant to adult clinical populations and to linguistic phenomena associated with neurologically based language disorders such as post-stroke aphasia, mild cognitive impairment, primary progressive aphasia, Parkinson’s disease, and Alzheimer’s disease. Symbols dedicated to child language development or to conversational analysis notation irrelevant to adult neurological speech were excluded from consideration from the start. The outcome of that evaluation was decisive: 40 symbols were retained unchanged, 15 were removed, four were adapted for Italian, and one was newly created. A parallel process addressed the 11 error-coding symbols, of which six were retained, four were adapted, and one new marker was introduced, with no symbols eliminated outright.

Many of the removals follow directly from the architecture of Italian. As a syllable-timed language, Italian does not systematically encode secondary stress at the lexical level, so the secondary stress marker was dropped. Vowel lengthening, common in Italian and heavily influenced by non-pathological dialectal variation, would confound clinical interpretation, so the lengthened-syllable marker went as well. Punctuation symbols such as the comma, semicolon, and colon were excluded because they neither define units of analysis nor influence automated CLAN outputs such as mean utterance length. The multiple-repetition marker was deemed redundant, already captured by the existing repetition symbol, and the unclear retracing notation was removed because it served only conversion to a competing transcription system. These are small deletions, but together they strip the transcription of artifacts that would otherwise inject noise into quantitative measures of Italian speech.

The adaptations cut deeper. In English, the omission of subjects, auxiliaries, or function words is typically a grammatical violation, so the original CHAT system codes such omissions with an error marker. Italian, however, is a null-subject language: dropping the pronoun in a phrase like “ha gli occhiali da sole”—literally “has the sunglasses”—is perfectly grammatical. The Italian toolkit therefore restricts the omitted-word marker to contexts where omission genuinely violates Italian grammar, and limits the grammatical error marker to utterances whose well-formedness fails by Italian, not English, standards. The clause delimiter was likewise repurposed to tag subordinate clauses, giving clinicians a direct metric of syntactic complexity that the English system computes automatically but that had no Italian equivalent—critical for characterizing the telegraphic, agrammatic output of many aphasic patients.

Phonological error coding underwent a deliberate simplification. The extensive English taxonomy of phonological error subtypes was collapsed into two operational categories: one for recognizable phonological errors such as omissions, substitutions, insertions, and metatheses, and one for phonological neologisms—complete losses of the intended target that may co-occur with non-word transcription symbols. Morphological coding was similarly tailored to the Italian inflectional system, split into markers for agreement errors of gender, number, and person, and for errors of verbal morphology including tense, mood, and irregular inflections. Because Italian verbs carry their grammatical freight in rich inflectional endings rather than in separate auxiliary words, morphological errors reveal a different facet of aphasia in Italian than they do in English.

Two entirely new symbols fill gaps that became glaring in clinical practice. The first, a word fragment marker, captures partial word productions that are abandoned without reformulation and without a precisely recognizable lexical target—a phenomenon the authors note is frequent in adult neurological speech and central to the lexical retrieval difficulties that constitute the most common aphasic disorder. The second, an articulatory distortion marker, codes phonetic distortions in which sounds are mangled despite correct phonological sequencing, a hallmark of motor speech impairments such as apraxia of speech and dysarthria. The original English system could represent such behaviors structurally, but could not quantify them at the error-coding level, even though distinguishing dysarthria from apraxia of speech from phonological difficulty is essential for differential diagnosis in the clinic.

To demonstrate the toolkit’s discriminative power, the researchers applied it to a connected speech sample from a 64-year-old right-handed woman with the non-fluent variant of primary progressive aphasia, a condition defined by effortful, agrammatic, often barely intelligible speech. Her sample, elicited with the picture-description task of the Screening for Aphasia in Neurodegeneration, was transcribed in Italian with the adapted conventions and compared with its English translation transcribed under the original CHAT rules. The side-by-side comparison exposes how the same clinical behavior is encoded differently depending on the language of analysis: the Italian transcript flags articulatory distortions, repeated phonological fragments, and retracing sequences through the new markers, while the English transcript must insert a null subject and assign a grammatical error code to subject omission that is simply licensed grammar in Italian. Compound nouns provide a further illustration, since English compounds rely on hierarchical stress patterns that Italian lexical items, carrying a single primary stress, do not.

Beyond the immediate benefit to Italian clinicians and researchers, the study stakes out a broader claim about how speech analysis should travel across languages. The authors emphasize that the majority of the world’s population does not speak English, yet research and clinical tools remain overwhelmingly Anglophone, and directly transferring English instruments often fails for diagnosis, treatment, and cross-cultural comparison. Their prior cross-linguistic work comparing healthy speakers of English, Italian, Cantonese, and Mandarin found significant differences in function word use and phonology that reflect language-specific morphosyntactic rules, reinforcing the point that linguistic diversity shapes the very manifestation of language processing patterns. The Italian toolkit therefore aims at a balance: preserving the conceptual organization and operational principles of TalkBank so that data remain comparable across languages, while introducing targeted adjustments grounded in the structure of Italian. It is designed as a flexible resource in which users select coding options guided by specific clinical or research goals, whether at the word or sentence level, or across the phonetic–phonological, lexical–semantic, morphosyntactic, and pragmatic domains.

The authors are candid about limitations. The adaptation targets adult speakers with neurological disorders and does not address developmental populations, where further modifications would be needed; coding remains time-consuming and may constrain large-scale implementation; and the morphosyntactic annotation layer—the %mor and %pos tiers—was deliberately left untouched, since adapting it requires alignment with both Italian grammar and specific research hypotheses. Future work, they suggest, should explore tighter integration between standardized coding schemes and automated transcription approaches such as Whisper-based speech recognition. Still, the toolkit stands as a replicable model for extending standardized annotation systems to any structurally distinct language, a necessary step if the rapidly advancing field of speech-based neurological biomarkers is to serve patients in the languages they actually speak.

Subject of Research: Cross-linguistic adaptation of standardized speech transcription and coding tools for assessing neurological language disorders in Italian-speaking adults.

Article Title: Extending standardized speech analysis across languages: An Italian toolkit for transcription and coding of neurological speech

Article References: Benzini, A., Lanzi, A. M., Palmirotta, C., Aresta, S., Battista, P., & Santi, G. C. (2026). Extending standardized speech analysis across languages: An Italian toolkit for transcription and coding of neurological speech. Behavior Research Methods, 58(10), Article 292. https://doi.org/10.3758/s13428-026-03175-x

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03175-x

Keywords: speech analysis, CHAT, CLAN, TalkBank, aphasia, primary progressive aphasia, cross-linguistic adaptation, Italian language, neurodegenerative disease, connected speech, language annotation, behavior research methods