Iyer Ravi, Nedeljkovic Maja, Meyer Denny
Centre for Mental Health, Swinburne University of Technology, Hawthorn, Australia.
JMIR Ment Health. 2022 Aug 15;9(8):e39807. doi: 10.2196/39807.
Artificial intelligence has the potential to innovate current practices used to detect the imminent risk of suicide and to address shortcomings in traditional assessment methods.
In this paper, we sought to automatically classify short segments (40 milliseconds) of speech according to low versus imminent risk of suicide in a large number (n=281) of telephone calls made to 2 telehealth counselling services in Australia.
A total of 281 help line telephone call recordings sourced from On The Line, Australia (n=266, 94.7%) and 000 Emergency services, Canberra (n=15, 5.3%) were included in this study. Imminent risk of suicide was coded for when callers affirmed intent, plan, and the availability of means; level of risk was assessed by the responding counsellor and reassessed by a team of clinical researchers using the Columbia Suicide Severity Rating Scale (=5/6). Low risk of suicide was coded for in an absence of intent, plan, and means and via Columbia suicide Severity Scale Ratings (=1/2). Preprocessing involved normalization and pre-emphasis of voice signals, while voice biometrics were extracted using the statistical language r. Candidate predictors were identified using Lasso regression. Each voice biomarker was assessed as a predictor of suicide risk using a generalized additive mixed effects model with splines to account for nonlinearity. Finally, a component-wise gradient boosting model was used to classify each call recording based on precoded suicide risk ratings.
A total of 77 imminent-risk calls were compared with 204 low-risk calls. Moreover, 36 voice biomarkers were extracted from each speech frame. Caller sex was a significant moderating factor (β=-.84, 95% CI -0.85, -0.84; t=6.59, P<.001). Candidate biomarkers were reduced to 11 primary markers, with distinct models developed for men and women. Using leave-one-out cross-validation, ensuring that the speech frames of no single caller featured in both training and test data sets simultaneously, an area under the precision or recall curve of 0.985 was achieved (95% CI 0.97, 1.0). The gamboost classification model correctly classified 469,332/470,032 (99.85%) speech frames.
This study demonstrates an objective, efficient, and economical assessment of imminent suicide risk in an ecologically valid setting with potential applications to real-time assessment and response.
Australian New Zealand Clinical Trials Registry ACTRN12622000486729; https://www.anzctr.org.au/ACTRN12622000486729.aspx.
人工智能有潜力革新当前用于检测自杀迫在眉睫风险的做法，并弥补传统评估方法的不足。
在本文中，我们试图根据澳大利亚两家远程医疗咨询服务机构接到的大量（n = 281）电话中自杀风险低与迫在眉睫的情况，对短语音片段（40毫秒）进行自动分类。
本研究纳入了总共281条求助热线电话录音，其中来自澳大利亚“在线”（On The Line）的有266条（94.7%），来自堪培拉000紧急服务中心的有15条（5.3%）。当来电者确认有意图、计划和手段时，将其编码为自杀迫在眉睫的风险；风险水平由接听的咨询师评估，并由一组临床研究人员使用哥伦比亚自杀严重程度评定量表（=5/6）重新评估。自杀风险低的情况是在没有意图、计划和手段以及通过哥伦比亚自杀严重程度量表评定（=1/2）时进行编码。预处理包括语音信号的归一化和预加重，同时使用统计语言R提取语音生物特征。使用套索回归识别候选预测因子。使用带有样条的广义相加混合效应模型评估每个语音生物标志物作为自杀风险的预测因子，以考虑非线性。最后，使用按分量梯度提升模型根据预编码的自杀风险评级对每个电话录音进行分类。
总共将77个高风险电话与204个低风险电话进行了比较。此外，从每个语音帧中提取了36个语音生物标志物。来电者性别是一个显著的调节因素（β = -0.84，95%置信区间 -0.85，-0.84；t = 6.59，P <.001）。候选生物标志物减少到11个主要标志物，并为男性和女性开发了不同的模型。使用留一法交叉验证，确保没有单个来电者的语音帧同时出现在训练和测试数据集中，精确率或召回率曲线下面积达到0.985（95%置信区间0.97，1.0）。gamboost分类模型正确分类了469,332/470,032（99.85%）个语音帧。
本研究展示了在生态有效环境中对自杀迫在眉睫风险进行客观、高效且经济的评估，具有在实时评估和应对中的潜在应用价值。
澳大利亚新西兰临床试验注册中心ACTRN12622000486729；https://www.anzctr.org.au/ACTRN12622000486729.aspx。
Zotero 插件