海洛因是一種容易高度成癮的毒品，並且會對社會產生相當大的負擔，也因此近年它得到了更多的重視。如何去分辨海洛因使用疾患(HUD)的嚴重程度以及給予他們適當的治療變成一個非常重要的議題。虛擬實境心率生理回饋(VR HRV biofeedback)訓練過去已被證明能夠有效改善毒癮患者的病情，這篇研究中，我們便採用此系統去對HUD患者進行治療，並同時收集他們的生理資訊，包括心電、皮膚電及呼吸訊號。 在這次研究中，我們收集了11位毒癮患者的六次訓練資料當作實驗組，以及10位正常人的一次訓練資料當作對照組，並利用統計以及機器學習的方式進行不同程度的患者生理訊號之分析。統計結果中，可以看到訓練前後比較起來，患者比起正常人有更多的顯著差異指標。而在機器學習結果可以發現，病人第一次訓練與正常人的資料進行分類可以達到86%的準確率，而病人第六次與正常人的分類結果變得較差，只有0.81的準確率。 此外，我們也採用了殘差網路(ResNet)對前述的分組進行分類，為了確認其直接用機器學習找出的時域資訊，是否能達到如同前面機器學習的結果，而結果中，顯示ResNet在分類患者第一次訓練以及正常人時，有達到接近的準確率。 ;Heroin is the highly addictive drug which produce numerous burdens to the society, so it became more and more important to distinct the different level of Heroin use disorder patients and give them proper treatment. The VR HRV biofeedback training, which was proved that it is enable to improve the HUD patients, was adopted as a treatment in this study, and the physiological signals, including electrocardiogram, galvanic skin response, and respiration, would be recorded during it. In this study, 11 HUD patients and 10 normal controls were invited to receive the biofeedback training, and the HUD patients would have 6 times of training. Through statistics and machine learning analysis of the signals, the differences of the signals of patient with different severity could be found. In the statistics analysis results, the HUD patient data had much more significant differences between pre- and post-test than the one of normal controls, which means the training affected the patients more than the normal controls. The machine learning results showed us the performance of classified pre-test of different group were better than the post-test, which means the patients became similar with normal controls after the training. Moreover, the results also presented the accuracies of classified HUD’s first training with the normal controls’ are higher than the one of HUD’s 6th training with normal controls. The ResNet methods were also employed in this study to find if it can have the better performance when only the time domain data be used as the input of model. It could reach the accuracies for 0.86 to classified the HUD patients and normal controls.
