Karlsson T
Department of Forensic Medicine, Karolinska Institute, Stockholm, Sweden.
Forensic Sci Int. 1998 Jun 22;94(3):183-200. doi: 10.1016/s0379-0738(98)00065-6.
Multivariate projective statistical methods (PCA, PLS-DA) and logistic regression analysis were used to create models to make predictions regarding whether a certain fatality shows similarities to homicide or suicide. The 'model set' consisted of 174 deaths due to sharp force injuries that in previous medicolegal investigation had been judged as homicides and 105 as suicides. The models were then validated on a new set of 40 homicides and 27 suicides that had not been used to create the models (test set validation). The model based on the PLS-DA technique had regarding its ability to identify homicides a sensitivity of 40/40 = 100% and a specificity of 25/27 = 93%. The model's predictions agreed with previously performed medicolegal investigations except in two suicides which according to the model were likely to be homicides. The reliability of this model was somewhat better than predictions achieved by means of logistic regression analysis, where six otherwise proven homicides were wrongly classified as suicides and two actual suicides were misclassified as homicides. The technique not only identifies variables but also ranks their importance. Ranked according to falling positive correlation (falling 'importance' of a finding) to the dependent variable 'death caused by homicide', the predictors were: Injuries to clothing, blood alcohol level, presence of defence injuries, injuries due to other type of violence than sharp force, chest stabs with vertical axis of the entrance wound, sharp force injuries to the upper extremity (except wrist and crook of the arm), sharp force injuries to the head and back. Ranked in increasing positive correlation to 'death caused by suicide' were the predictors: sharp force injuries to the crook of the arm, venue being the victim's home, presence of farewell letter, victim's age, sharp force injuries to the wrist, known suicidal ideation and presence of tentative injuries.
使用多元投影统计方法（主成分分析、偏最小二乘判别分析）和逻辑回归分析来创建模型，以预测某一死亡事件是否与他杀或自杀具有相似性。“模型集”包括174例因锐器伤致死的案例，在之前的法医学调查中被判定为他杀，以及105例被判定为自杀。然后，在一组未用于创建模型的40例他杀和27例自杀案例（测试集验证）上对模型进行验证。基于偏最小二乘判别分析技术的模型在识别他杀方面的敏感性为40/40 = 100%，特异性为25/27 = 93%。该模型的预测结果与之前进行的法医学调查结果一致，但有两例自杀案例根据模型可能被判定为他杀。该模型的可靠性略优于通过逻辑回归分析获得的预测结果，在逻辑回归分析中，有6例经证实的他杀案例被错误地分类为自杀，2例实际自杀案例被错误地分类为他杀。该技术不仅能识别变量，还能对其重要性进行排序。根据与因他杀导致的死亡这一因变量的正相关性下降（发现的“重要性”下降）进行排序，预测因素为：衣物损伤、血液酒精水平、防御伤的存在、除锐器伤之外其他类型暴力导致的损伤、入口伤口垂直轴的胸部刺伤、上肢（手腕和肘部除外）的锐器伤、头部和背部的锐器伤。与因自杀导致的死亡的正相关性增加的预测因素排序为：肘部的锐器伤、案发地点为受害者家中、遗书的存在、受害者年龄、手腕的锐器伤、已知的自杀意念和试探性损伤的存在。
