Psychosocial health issues pose a significant global health crisis, contributing substantially to the global burden of disease and economic costs. The COVID-19 pandemic exacerbated this challenge, leading to a marked increase in stress, anxiety, and depression worldwide. Family caregivers, a growing population due to increased longevity and chronic illnesses, are particularly vulnerable to psychosocial health problems due to the substantial physical, emotional, and financial burdens of caregiving. Early detection and intervention are crucial for preventing disease progression and reducing long-term disability. However, traditional psychosocial health assessments, typically involving lengthy interviews with specialists, are cumbersome, subjective, and often inaccessible. This study addresses these limitations by developing and evaluating ASAP, a digital tool using speech analysis to assess stress levels in family caregivers. The program leverages natural language processing (NLP) techniques to analyze speech patterns and identify keywords indicative of stress, offering a more efficient and potentially less biased alternative to traditional methods.

Existing research has explored the use of NLP techniques to detect mental illness from social media posts, but this approach excludes non-netizens. Furthermore, these methods often lack explanatory power regarding the causes of mental illness. This study builds upon previous findings that analyzing word frequency in spoken text can indicate various psychosocial disorders. The researchers propose using topic modeling to discover underlying themes in speech, recognizing that both conscious and unconscious elements of speech reflect psychosocial status.

The ASAP workflow consists of five stages: (1) Automated Speech Recognition (ASR) using the Google Cloud Speech API to transcribe Cantonese speech into text; (2) Text Pre-processing (TP) using the PyCantonese Python package for word segmentation and stop word removal; (3) Term Frequency-Inverse Document Frequency (TF-IDF) calculation to identify significant words; (4) Text Analytics (TA) employing a topic ensemble method (combining Non-negative Matrix Factorization and hierarchical clustering) to discover abstract topics in the speech; and (5) Visual Analysis (VA) to display clustering results. The study involved 100 Cantonese-speaking family caregivers recruited from a non-profit organization. Participants were screened for caregiver stress burden using the Caregiver Burden Inventory (CBI), a 24-item self-report scale. Caregivers were then interviewed, answering twelve open-ended questions related to Walsh's family resilience theory. The interviews were recorded and analyzed using the ASAP. The performance of the proposed unsupervised learning method in ASAP was compared with five supervised learning methods: three versions of Support Vector Machine (SVM) and two deep learning methods (Word Embedding and Recurrent Neural Network). A 10-fold cross-validation technique was used for the supervised learning methods.

The ASAP successfully distinguished between family caregivers with high and low stress burden levels. The analysis grouped caregivers into two clusters (A and B). Cluster A contained 38 of the 53 caregivers with low stress levels, while Cluster B contained 34 of the 47 caregivers with high stress levels, resulting in a 72% accuracy rate. This accuracy surpassed that of five other machine learning models tested. The ASAP's performance was also found to be robust, maintaining high accuracy even when only a subset of words (60%) were used. The keywords identified in each cluster differed significantly, with low-stress caregivers exhibiting topics related to relaxation and positive emotions (e.g., travel, gratitude) and high-stress caregivers demonstrating topics focused on family issues and negative emotions (e.g., concerns about family members, lack of satisfaction in relationships). The average interview time was approximately 30 minutes, significantly shorter than traditional assessment methods.

The ASAP demonstrates the feasibility of using digital health technology for efficient and accurate assessment of stress burden in family caregivers. The high accuracy rate (72%) and the efficiency of the process (30-minute interviews compared to 1-2 hour traditional assessments) highlight the potential of this tool to improve access to timely psychosocial health interventions. The identification of distinct keywords and topics in the two clusters provides insights into the verbal expressions associated with high and low stress levels. The study suggests that analyzing responses to non-sensitive questions related to family resilience can reduce bias in psychosocial health assessments. The objective nature of the machine learning approach provides consistent and reliable assessments, while the program’s scalability offers potential for wider application in the future.

The ASAP offers a valuable tool for early detection and intervention in stress and psychosocial health issues among family caregivers. Its efficiency and accuracy make it a promising candidate for integration into healthcare systems to address the growing global psychosocial health crisis. Further research should focus on improving accuracy, optimizing the number and content of interview questions, and extending ASAP's capabilities to detect a wider range of psychosocial problems.

The study's limitations include the use of a clear-cut threshold (score of 36) on the CBI to define low versus high stress, which might affect accuracy. Further research is needed to validate this threshold and enhance classification accuracy. Also, the current study's findings may not be generalizable to populations beyond Cantonese-speaking family caregivers. The limited sample size (100 participants) could also influence the generalizability of the results. The integration of digital health programs into healthcare systems remains a challenge, requiring attention to both technological and organizational factors.