Sunghye Cho

Purpose: Examine how task structure and participant sex influence social-word production in verbally fluent autistic and non-autistic youth during two ADOS‑2 Module 3 tasks (Picture Description and Telling a Story From a Book). Methods: One hundred eleven verbally fluent autistic and non-autistic youth (ages 6–15) completed ADOS‑2 Module 3. Groups were matched on age and IQ. Transcripts of Picture Description and Telling a Story From a Book were processed with LIWC‑22 to quantify social words. Generalized linear models predicted social word production in each task using age, IQ, sex, and diagnostic group as predictors, including sex-by-diagnosis interactions. Associations between social-word use in each task and ADOS‑2 Social Affect scores were examined within the autistic group. Results: In the Picture Description task, there was a significant interactive effect of sex and diagnosis. This was driven by non-autistic girls using more social words than any other subgroup. Autistic girls and boys did not differ. In the Telling a Story From a Book task, there were no significant effects of sex or diagnosis. Within the autistic group, participants who used fewer social words during Picture Description task were rated as having greater social difficulties by expert clinicians. Conclusion: Task structure impacts detection of diagnostic and sex-related differences in social language: less structured contexts reveal diagnostic group differences, while structured narrative elicitation attenuates them. Social word production in unstructured settings relates to clinician-rated social difficulties, particularly for autistic girls, highlighting the importance of task selection when assessing pragmatic language.

Development and Validation of the 9-item Parkinson’s Disease Screening Questionnaire

Autistic individuals often struggle with social communication, but the nature of each person’s difficulties varies substantially by context. Recent conversation-based research suggests that males and females (autistic and nonautistic) produce subtle but consistent differences in language during spontaneous conversation. However, the majority of prior studies of conversation in autism include female conversation partners only rather than systematically manipulating whether conversation dyads were same-sex or mixed-sex, thus limiting the generalizability of the findings to everyday life. In this first-of-its-kind study, we examine speech-language behavior in 35 autistic adolescents engaged in two 5-min conversations with nonautistic partners who were sex-matched or -mismatched to the participant. Data were collected at a public research university in the United States as part of a baseline assessment for a randomized control trial that concluded in 2015. Linear mixed effects regression revealed significant effects of sex-match or -mismatch on language, such that participants in same-sex conversations were more talkative, took longer turns, and paused differently than participants in mixed-sex conversations. Results highlight the importance of considering a partner’s demographic characteristics when designing studies involving natural language in autism and suggest that prior research with single-sex conversation partners should be interpreted cautiously as results may not generalize to real-world contexts. Improved understanding of everyday conversational language in autism can be used to inform neurodiversity-affirming supportive interventions while also identifying robust linguistic markers of social phenotype that can be used as outcome measures in future clinical trials. Lay Autistic individuals often speak and interact differently than nonautistic people, which can lead to challenges when these groups engage in conversation together. Communication can look very different according to the situation, however, including who is in the conversation. For example, prior research found sex differences in autistic language during natural, get-to-know-you conversations, but most existing studies did not consider the potential influence of who a person is talking to. In other words, does it matter whether you are talking to someone that is similar to you versus different from you? This gap in prior research limits our understanding of how natural conversation plays out in real life, when individuals talk to a wide variety of people. In this first-of-its-kind study, we examine whether autistic participants’ language changes depending on whether they talk to a male or a female conversation partner. Thirty-five female and male autistic adolescents participated in two 5-min conversations with nonautistic partners who were sex-matched or -mismatched to the participant. Analyses showed that participants in same-sex conversations were more talkative, took longer turns, and paused differently than participants in mixed-sex conversations. Results highlight the importance of considering partner characteristics when analyzing natural language in autism and suggest that prior research should be interpreted cautiously as results may not generalize to real-world contexts. This study contributes to the goal of better understanding conversational language in autism.

Purpose: The rapid advancement of automatic speech recognition (ASR) and natural language processing technologies has created significant opportunities for clinical applications within speech and language disorders, yet these capabilities remain largely confined to high-resource languages and populations. As research communities work to address these inequities through inclusive speech data collection, the intersection of clinical vulnerability, linguistic diversity, and emerging speech and language technologies creates ethical considerations that are rarely addressed by existing guidelines. Ethical data collection practices affect the fairness and bias profiles of automatic speech and language analysis systems trained on these data, creating a foundational link between participant protection and algorithmic justice. Method: This article introduces the Protected Entities Ethics Checklist (PEEC), a comprehensive framework specifically designed for researchers collecting speech and language data from populations requiring enhanced protections. The framework addresses three core domains: participant protection and consent, data collection standards, and compliance implementation. Critically, the PEEC situates ethical data collection as a prerequisite for developing fair ASR systems, recognizing that procedural justice in research must precede algorithmic fairness. Results: The PEEC framework provides structured guidance for ethical research with protected entities including children, elderly adults with cognitive changes, individuals with communication disorders, and marginalized communities. It offers population-specific consent mechanisms, enhanced data protection measures, systematic quality assurance procedures, and explicit guidance on technical considerations for ASR applications while maintaining flexibility for diverse research contexts. Conclusions: Ethical treatment of research participants is inextricably linked to algorithmic fairness in speech technology development. The PEEC framework argues that procedural justice in data collection is a prerequisite for achieving fair AI systems, establishing the necessary ethical foundation for subsequent technological development in clinical speech research. By ensuring equitable and respectful data collection practices, we create the foundation for ASR systems that perform equitably across diverse populations.

Background: Early clinical symptoms of both Alzheimer's disease (AD) and dementia with Lewy body (DLB) include memory problems and difficulties in judgement and planning. With similar clinical symptoms, it is challenging to screen patients with AD at scale while excluding those with DLB. This study proposes automatic classification systems that utilize both text and acoustic features to automatically distinguish patients with AD from those with DLB. Method: Patients with AD pathology, including both AD and mild cognitive impairment (MCI) due to AD (ADCI; n = 55, 35 females, age=75.69±6.79y, education=9.78±4.49y), those with DLB or MCI due to DLB (DLBCI; n = 28, 11 females, age=77.93±5.79y, education=9.14±4.49y), and healthy elderly speakers (HC; n = 55, 29 females, age=72.51±6.35y, education=9.87±4.19y) completed a 5‐minute speech collection protocol that included one picture description task, two semantic fluency tasks, three letter fluency tasks, and one sustained phonation task. Amyloid positivity was confirmed with visual assessment of 18F‐Florbetaben PET, and all patients in the ADCI group exhibited amyloid positivity. We employed a state‐of‐the‐art large language model (KLUE‐RoBERTa‐base) to extract token‐level embeddings from concatenated texts of 1 picture description and five fluency tasks, which were mean‐pooled to generate participant‐level embedding features ( n = 768). OpenSMILE (the EMOBASE configuration) was employed to extract acoustic features ( n = 988) from the sustained phonation task. We applied Principal Component Analysis to reduce feature dimensions in each feature set respectively, and we experimented with two fusion methods (data‐level vs. decision‐level) in training Multilayer Perceptron classifiers for two binary tasks: HC vs. all patients (ADCI+DLBCI) and ADCI vs. DLBCI. The models were validated using five‐fold cross validation. Result: Table 1 summarizes all results. The classifier for distinguishing HC from all patients showed the best performance (accuracy=75.7%, AUC=0.77) when the text embeddings and the acoustic features were fused at the feature level. Text‐only or acoustic‐only models showed lower performances compared to the fused models. For distinguishing between ADCI and DLBCI, the model trained with decision‐level fusion showed the best performance (accuracy=80.3%, AUC=0.82), followed by the feature‐level fusion model (accuracy=77.8%, AUC=0.79). Conclusion: Our study demonstrates that the automatic classification of ADCI and DLBCI using speech is highly promising for large‐scale screening of patients for clinical trials and interventions.

Automated systems that analyze short conversation samples could offer valuable insights into autistic children's real-world social communication competence while highlighting opportunities for growth. In this study, we explore the potential of using large language models (LLMs) to assess social interaction quality during casual conversations that include autistic youth. We focus on evaluating the ability of LLMs to predict conversation quality, with the long-term goal of informing tailored communication support strategies that promote optimal social outcomes. Results underscore the need for caution when using LLMs to inform clinical decision-making in autism.

Background: Story recall tasks are often employed in clinical settings to measure verbal episodic memory in individuals with mild cognitive impairment (MCI) and Alzheimer’s disease (AD). Traditional analyses (e.g. total number of recalled words) provide useful but limited information compared to detailed analyses of lexical performance. In this study, we analyzed story recall data produced by patients with AD or MCI using an automated method.Method: We analyzed digitized recordings of the Craft story recall task produced by 15 individuals with MCI (age=65.8±10.14, females=40%, education=16.07±2.94), 27 individuals with AD (age=65.19±6.82, females=37%, education=15.42±3.01) and 14 healthy age‐matched individuals (HC; age=58.86±16.16, females=50%, education=17.29±2.16). The groups did not differ in any demographic characteristics. Our automated pipeline identified the number of verbatim and paraphrase recalls, calculated semantic similarities between the original story and recalled immediate and delayed stories using word embeddings and rated several lexical characteristics, including word frequency, concreteness, and word length. Two‐way ANOVA tests examined the effect of Group (HC vs. MCI vs. AD), Task (immediate vs. delayed), and their interactions. As the interactions of Group and Task were not significant, post‐hoc analyses examined significant pairwise group differences in immediate and delayed recalls, separately.Result: Participants with AD produced the fewest verbatim and total recalled words, followed by MCI patients and HC in both immediate and delayed recalls (Figure 1A–B, Table 1). AD patients produced fewer paraphrase recalls compared to the other groups in the delayed recall only (Figure 1C). AD group’s recalled story was also the least similar to the original story, followed by MCI and HC (Figure 1D). Also, individuals with AD produced words that were more frequent and less concrete during immediate and delayed recalls compared to MCI patients and HC (Figure 1E–F). Words produced by AD patients were also shorter than those produced by the MCI and HC groups during the immediate recall (Figure 1G).Conclusion: Our method highlights that digitized language‐based tests of verbal episodic memory can be automatically analyzed. These capture traditional language features that cannot be extracted manually without massive effort and promise to provide novel, granular, digitized biomarkers that traditional assessment methods of neuropsychological tests cannot show.

INTRODUCTION: Screening for Alzheimer's disease neuropathologic change (ADNC) in individuals with atypical presentations is challenging but essential for clinical management. We trained automatic speech‐based classifiers to distinguish frontotemporal dementia (FTD) patients with ADNC from those with frontotemporal lobar degeneration (FTLD).METHODS: We trained automatic classifiers with 99 speech features from 1 minute speech samples of 179 participants (ADNC = 36, FTLD = 60, healthy controls [HC] = 89). Patients’ pathology was assigned based on autopsy or cerebrospinal fluid analytes. Structural network‐based magnetic resonance imaging analyses identified anatomical correlates of distinct speech features.RESULTS: Our classifier showed 0.88 0.03 area under the curve (AUC) for ADNC versus FTLD and 0.93 0.04 AUC for patients versus HC. Noun frequency and pause rate correlated with gray matter volume loss in the limbic and salience networks, respectively.DISCUSSION: Brief naturalistic speech samples can be used for screening FTD patients for underlying ADNC in vivo. This work supports the future development of digital assessment tools for FTD.Highlights: We trained machine learning classifiers for frontotemporal dementia patients using natural speech. We grouped participants by neuropathological diagnosis (autopsy) or cerebrospinal fluid biomarkers. Classifiers well distinguished underlying pathology (Alzheimer's disease vs. frontotemporal lobar degeneration) in patients. We identified important features through an explainable artificial intelligence approach. This work lays the groundwork for a speech‐based neuropathology screening tool.