This study developed an AI model that can distinguish individuals who died from suicide using publicly available web-based real-world audio data. The most effective model, the feedforward neural network (multilayer persepron), demonstrated high prediction accuracy, particularly considering accidental reliance on real-world data. In particular, this accuracy was significantly improved when analyzing a subset of near group data, including individuals who died from suicide within 12 months of audio recording. Specifically, this model increased the AUC and accuracy of this subset, highlighting the important role of temporal proximity in the identification of speech biomarkers associated with suicide risk.
These findings demonstrate previous studies suggesting that Para-Ringo traits are promising biomarkers for mental health conditions, including suicide. For example, previous work by Pestian et al.9,10 and Hashim et al.11,12 We demonstrated that acoustic and linguistic features can distinguish individuals with suicidal ideation from control with high accuracy. However, these studies relied on clinical or structured settings rather than actual suicide, predicting suicidal ideation or questionnaire results.
Our work builds and expands these efforts by demonstrating that AI models can predict completed suicides using publicly available naturalistic speech data. This is Amiriparian et al.13 and Song et al.14Although highly accurate, it relies on emergency or hotline data and focuses on risk indicators rather than confirmed results. By leveraging data from the general population and focusing on validated suicide deaths, the model is more closely aligned with the real application possibilities, and Iyer et al.17 and Belouali et al.18.
Furthermore, our findings are Walsh et al. reflects observations by34Who found that temporary proximity to an event would improve suicide risk prediction? It shows a similar trend. Prediction accuracy increases significantly if suicide occurs within 12 months of recording, suggesting that acoustic markers of suicide risk may become more pronounced as the event approaches.
Furthermore, the model achieved strong results not only on the Feedforward Neural Network, but also across multiple classification algorithms such as logistic regression, nearest neighbors, Xgboost linearity, and Xgboost tree. The robustness of this study is further demonstrated by the ability of the model to maintain high performance despite perturbations in the analytical pipeline. These findings suggest the potential for significant advances in suicide risk assessment through the use of ML models, particularly neural network-voice data.
Overall, despite the inherent complexity and limitations of the data, the model exhibits significant robustness, which is characterized by noise, lack of background clinical information, which may shed light on the potential causes of suicide completed, and the difficulty of the task at hand. However, it is important to note that in our study there is no clinical information detailing the specific psychopathology experienced by participants by those who died from suicide and those in the control group. Despite the need to make assumptions about the external effects of potential comorbidities and drugs of abuse, the dataset remains invaluable due to its clear “hard outcomes.” However, this inherent drawback can be viewed as a strength of our research. That is, we were able to identify clear groups, even amidst the lack of background clinical information.
Importantly, this is the first study to successfully predict actual suicidal behavior, rather than relying on surrogate markers, such as self-report measures from surveys. This illustrates important advances in suicide prevention research to demonstrate the possibility of demonstrating the possibility of using AI to analyze naturalistic speech data for identification of suicide risk. These findings promise to improve early detection efforts and coordinate interventions to prevent suicide, particularly at critical times leading up to the event.
Future research
Our future research efforts will focus on rigorous verification of results by acquiring data in clinical studies featuring well-diagnosed patients. Furthermore, investigations of additional acoustic functional groups, such as subtle analyses of tempo, rhythm, spectral features, and pauses (both fulfilled and reclaimed) present a means to further improve model performance. It allows for consideration of larger datasets based on automated analysis, holdout sets, and improves current cross-validation approaches for a more robust assessment of results. Furthermore, by incorporating a broader dataset, it may facilitate the application of alternative end-to-end AI approaches, including trans models. Furthermore, textual information (using a particular word class or a specific ratio of word class) is required to make the analysis more comprehensive. Furthermore, distinguishing between the various psychopathology that contributes to depression, suicidal ideation, and suicide death within the general population will allow us to gain a deeper understanding of the complexities surrounding mental health. One way to achieve these goals is to integrate speech biomarker analysis into the national suicide prevention hotline.
Ethical considerations
The integration of AI and ML in health data analysis poses many ethical issues to be carefully considered. It is essential to protect individual privacy while opposing the potential benefits of happiness. As emphasized by Lejeune et al.35 “The application of AI to health data requires robust cybersecurity and a clear legal framework” (p. 7). Researchers in this group further emphasize that AI has a huge commitment but requires caution as it can rely on issues of responsibility, namely technology in healthcare decisions. They argue that such trust could lead to reduced human surveillance and accountability in clinical settings. Therefore, AI should be viewed as a complementary tool rather than replacing human clinical judgment. Ensuring that healthcare professionals are crucial to the decision-making process is essential to maintaining quality and ethical standards of patient care. Furthermore, while this study highlights the possibility of reusing publicly available data to address important health challenges and gain valuable insights (data collection procedures verified by the Ethics Committee), we strongly agree that the development and deployment of AI systems in healthcare is guided by basic ethical principles, transparent and responsible responsibility.
