ZLZexu Li

Zexu Li

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zexu Li has a total of 36 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Automated Speech Features from Logical Memory Delayed Recall Test Relate to Early Tau PET Burden

    by Zexu Li, Christina B. Young, Cody Karjadi, Alexandra König, Louisa Schwed, Nicklas Linz, Johannes Tröger, Julia Peterson, Jinying Chen, Ting Fang Alvin Ang, et al.

    Background: Speech features (e.g., between‐utterance pause duration, speech rate) derived from Logical Memory Delay Recall (LMd) tests were shown to be associated with early tau burden. Various software tools can generate these features but may use different methods. We aim to validate the robustness of speech features by comparing analogous features generated by two software tools and assessing their associations with tau pathology. Method: We analyzed data from Framingham Heart Study participants who completed amyloid and tau PET imaging within a year of their LMd test. Five speech features were generated using ki:elements’ SIGMA platform and the CLAN software. Ki:elements’ tool derived three features (duration of utterances, duration of between‐utterance pauses, number of between‐utterance pauses) fully automatically from voice recordings. Other Ki:elements features and all CLAN features were derived from manual transcription. Correlations between ki:elements and CLAN features were assessed by Spearman's correlation coefficients. Associations between speech features and tau standardized uptake value ratio (SUVR) in five brain regions were examined using multiple linear regression, adjusting for age, sex, education, and amyloid status. Result: Data from 237 participants (mean age: 54.6±8.7 years; 51.1% female) were analyzed (Table 1). All pairs of speech features were correlated ( p p = 0.03;CLAN: beta=2.74, p = 0.002), inferior temporal (IT) (Ki:elements: beta=2.34, p = 0.03; CLAN: beta=2.34, p = 0.03), and inferior parietal (IP) (Ki:elements: beta=3.45, p p p = 0.02; CLAN: beta=‐2.74, p = 0.002) and IT (Ki:elements: beta=‐3.10, p = 0.006; CLAN: beta=‐2.51, p = 0.02) regions. Conclusion: Simple speech features (including fully automated ones) generated using two methods exhibited compatible patterns in their associations with early tau burden. Other automated linguistic and semantic features will be further investigated. With continued validation, automated speech features have the potential to offer a scalable approach with limited loss of findings relative to manual transcriptions.

  • Open Access

    Developing New Digital Markers for Mild Cognitive Impairment: A Large Language Model‐Based Approach

    by Zexu Li, Haochun Huang, Ting Fang Alvin Ang, Rhoda Au, Jinying Chen

    Background: The logical memory delayed (LMd) recall test has been widely used in screening tools for Alzheimer's disease (AD) and mild cognitive impairment (MCI). The test performance was evaluated by trained professionals based on the number of predefined key terms the participant accurately recalled from a previously narrated standard story. This process is subjective and labor‐intensive. We aim to develop novel, automated digital markers for MCI using large language models (LLMs) and transcriptions of recorded verbal responses from the LMd test. Method: We analyzed data from participants of the Framingham Heart Study who (1) had at least one audio transcription from the LMd test and (2) completed all the neuropsychological (NP) tests (in the same visit as (1)) used for assessing MCI based on the modified Jak/Bondi criteria. We developed automated measures that assessed participants’ LMd performance by using LLM‐derived text embeddings to estimate the semantic similarity between the standard LMd test story (72 words) and the transcriptions of verbal retellings (average 62 words). We implemented four measures using state‐of‐the‐art LLMs for short‐text embedding: E5, MiniLM, MPNet, and the Universal Sentence Encoder (USE). We evaluated these measures against the LMd score through two analyses. First, we used a generalized linear mixed model to assess the association between each measure and MCI status, adjusting for age, sex, and education. Second, we used logistic regression to assess the association between each measure at baseline and the conversion from cognitively normal to MCI within 15 years, adjusting for age, sex, education, and the conversion duration. Result: 587 LMd recordings from 282 FHS participants (age at baseline: 53.7±9.8, 48.9% Female) were analyzed, with 78 recordings corresponding to MCI cases (Table 1). The LLM‐derived measures were correlated with the LMd score (Figure 1). Most LLM‐derived measures were negatively associated with MCI status (Table 2), with USE‐derived measure showing the strongest association (OR=0.539, P < 0.001). All LLM‐derived measures were negatively associated with normal‐to‐MCI conversion (Table 2), with USE‐derived measure being most sensitive (OR=0.344, P < 0.001). Conclusion: With further validation, these novel LLM‐derived measures can potentially serve as digital markers for MCI.

  • Open Access

    Exploring the Perspectives of Older Adults on a Digital Brain Health Platform Using Natural Language Processing: Cohort Study

    by Huitong Ding, Katherine Gifford, Ludy C Shih, Kristi Ho, Salman Rahman, Akwaugo Igwe, Spencer Low, Zachary Popp, Edward Searls, Zexu Li, et al.

    Background: Although digital technology represents a growing field aiming to revolutionize early Alzheimer disease risk prediction and monitoring, the perspectives of older adults on an integrated digital brain health platform have not been investigated. Objective: This study aims to understand the perspectives of older adults on a digital brain health platform by conducting semistructured interviews and analyzing their transcriptions by natural language processing. Methods: The study included 28 participants from the Boston University Alzheimer’s Disease Research Center, all of whom engaged with a digital brain health platform over an initial assessment period of 14 days. Semistructured interviews were conducted to collect data on participants’ experiences with the digital brain health platform. The transcripts generated from these interviews were analyzed using natural language processing techniques. The frequency of positive and negative terms was evaluated through word count analysis. A sentiment analysis was used to measure the emotional tone and subjective perceptions of the participants toward the digital platform. Results: Word count analysis revealed a generally positive sentiment toward the digital platform, with “like,” “well,” and “good” being the most frequently mentioned positive terms. However, terms such as “problem” and “hard” indicated certain challenges faced by participants. Sentiment analysis showed a slightly positive attitude with a median polarity score of 0.13 (IQR 0.08-0.15), ranging from –1 (completely negative) to 1 (completely positive), and a median subjectivity score of 0.51 (IQR 0.47-0.53), ranging from 0 (completely objective) to 1 (completely subjective). These results suggested an overall positive attitude among the study cohort. Conclusions: The study highlights the importance of understanding older adults’ attitudes toward digital health platforms amid the comprehensive evolution of the digitalization era. Future research should focus on refining digital solutions to meet the specific needs of older adults, fostering a more personalized approach to brain health.

  • Open Access

    Association of DNA methylation age acceleration with digital clock drawing test performance: the Framingham Heart Study

    by Zexu Li, Huitong Ding, Mengyao Wang, Yi Li, Ting Fang Alvin Ang, Gurnani Ashita, Gifford Katherine, Cody Karjadi, Daniel Levy, Rhoda Au, et al.

    Background: Cognitive function measured by digital clock drawing test (dCDT) has drawn attention for their precision, automation, and reproductivity. However, the relationship between digital cognitive metrics and biological aging is lacking. Methods: We conducted association analyses between cognitive function measured by dCDT and biological aging metrics quantified by five DNA methylation (DNAm) age metrics (Horvath, Hannum, GrimAge, PhenoAge, and DunedinPACE) in the Framingham Heart Study (FHS). We conducted linear regression to investigate the association between cognitive functions (global cognitive function and four sub-domain functions) and DNAm age acceleration, adjusting for covariates. We used a false discovery rate (FDR) < 0.05 for significance. Results: Among the 1,798 FHS participants (mean age 65±13, 53% women), we found that a lower dCDT total score is associated with DNAm age acceleration. Larger magnitudes of associations were observed in older participants (≥ 65 years). The dCDT total score showed the strongest association with the DundinPACE in the pooled sample (beta = −2.1, FDR = 0.0004), the younger (beta = −1.9, FDR = 0.02), and older age group (beta = −2.2, FDR = 0.01). The dCDT total score was significantly associated with age acceleration estimated by Horvath (beta=-1.9, FDR =0.01) and PhenoAge (beta=-2.5, FDR=0.01) in older participants while not in the pooled sample or younger participants (<65 years). In sub-domain cognitive functions, we found that simple motor function was significantly associated with DunedinPACE (FDR = 0.005) in both age groups and associated with GrimAge (FDR = 0.05 in older age group), indicating the deterioration in various organ systems may particularly impact this domain. Discussion: Our findings suggest that cognitive function measured by a digital clock drawing test is associated with DNAm age acceleration in middle-aged and older participants in the FHS, potentially shedding light on the epigenetic mechanisms underlying digitally measured cognitive function.

  • Open Access

    Assessment of Wearable Device Adherence for Monitoring Physical Activity in Older Adults: Pilot Cohort Study

    by Huitong Ding, Kristi Ho, Edward Searls, Spencer Low, Zexu Li, Salman Rahman, Sanskruti Madan, Akwaugo Igwe, Zachary Popp, Alexa Burk, et al.

    Background: Physical activity has emerged as a modifiable behavioral factor to improve cognitive function. However, research on adherence to remote monitoring of physical activity in older adults is limited. Objective: This study aimed to assess adherence to remote monitoring of physical activity in older adults within a pilot cohort from objective user data, providing insights for the scalability of such monitoring approaches in larger, more comprehensive future studies. Methods: This study included 22 participants from the Boston University Alzheimer’s Disease Research Center Clinical Core. These participants opted into wearing the Verisense watch as part of their everyday routine during 14-day intervals every 3 months. Eighteen continuous physical activity measures were assessed. Adherence was quantified daily and cumulatively across the follow-up period. The coefficient of variation was used as a key metric to assess data consistency across participants over multiple days. Day-to-day variability was estimated by calculating intraclass correlation coefficients using a 2-way random-effects model for the baseline, second, and third days. Results: Adherence to the study on a daily basis outperformed cumulative adherence levels. The median proportion of adherence days (wearing time surpassed 90% of the day) stood at 92.1%, with an IQR spanning from 86.9% to 98.4%. However, at the cumulative level, 32% (7/22) of participants in this study exhibited lower adherence, with the device worn on fewer than 4 days within the requested initial 14-day period. Five physical activity measures have high variability for some participants. Consistent activity data for 4 physical activity measures might be attainable with just a 3-day period of device use. Conclusions: This study revealed that while older adults generally showed high daily adherence to the wearable device, consistent usage across consecutive days proved difficult. These findings underline the effectiveness of wearables in monitoring physical activity in older populations and emphasize the ongoing necessity to simplify usage protocols and enhance user engagement to guarantee the collection of precise and comprehensive data.


  • Open Access

    Exploring the Perspectives of Older Adults on a Digital Brain Health Platform Using Natural Language Processing: Cohort Study (Preprint)

    by Huitong Ding, Katherine Gifford, Ludy C Shih, Kristi Ho, Salman Rahman, Akwaugo Igwe, Spencer Low, Zachary Popp, Edward Searls, Zexu Li, et al.

    BACKGROUND Although digital technology represents a growing field aiming to revolutionize early Alzheimer disease risk prediction and monitoring, the perspectives of older adults on an integrated digital brain health platform have not been investigated. OBJECTIVE This study aims to understand the perspectives of older adults on a digital brain health platform by conducting semistructured interviews and analyzing their transcriptions by natural language processing. METHODS The study included 28 participants from the Boston University Alzheimer’s Disease Research Center, all of whom engaged with a digital brain health platform over an initial assessment period of 14 days. Semistructured interviews were conducted to collect data on participants’ experiences with the digital brain health platform. The transcripts generated from these interviews were analyzed using natural language processing techniques. The frequency of positive and negative terms was evaluated through word count analysis. A sentiment analysis was used to measure the emotional tone and subjective perceptions of the participants toward the digital platform. RESULTS Word count analysis revealed a generally positive sentiment toward the digital platform, with “like,” “well,” and “good” being the most frequently mentioned positive terms. However, terms such as “problem” and “hard” indicated certain challenges faced by participants. Sentiment analysis showed a slightly positive attitude with a median polarity score of 0.13 (IQR 0.08-0.15), ranging from –1 (completely negative) to 1 (completely positive), and a median subjectivity score of 0.51 (IQR 0.47-0.53), ranging from 0 (completely objective) to 1 (completely subjective). These results suggested an overall positive attitude among the study cohort. CONCLUSIONS The study highlights the importance of understanding older adults’ attitudes toward digital health platforms amid the comprehensive evolution of the digitalization era. Future research should focus on refining digital solutions to meet the specific needs of older adults, fostering a more personalized approach to brain health.