探討活動季節變化與情緒疾患及憂鬱狀態的關係

Abstract

重鬱症與雙極性情感疾患為常見且經常反覆發作的情緒疾患，憂鬱情緒為情緒疾患所共有的特徵。過去研究指出，患者的許多臨床症狀，被觀察到容易隨著季節的更迭而有所改變，其中活動模式以及睡眠狀況是日常中最容易觀察的表徵資訊，除了主觀的陳述之外，若有完整的客觀表現型數據，將能協助臨床醫療人員更快速判斷患者的診斷，或其當下疾病的嚴重程度，甚至未來能更進一步提前偵查症狀發作前的徵兆。故本研究主要探討活動模式於季節變化中與情緒疾患及憂鬱特徵的關係，我們將以腕動計所蒐集到的活動與睡眠表徵資訊，對情緒障礙的診斷做分類，並了解臨床症狀與實際活動量之間的相關性。 本篇研究招募49位雙極性情感疾患以及52位憂鬱症患者，且從社區招募41位過去無精神科就醫紀錄的健康受試者，進行四季各為期一周的追蹤研究，其中53位完成完整四季的追蹤，其餘的收集一到三季不等的資訊。我們使用楊氏躁症量表、漢氏憂鬱量表與貝氏憂鬱量表，評估症狀的嚴重程度。此研究也使用季節性問卷，評估情緒與季節變化之主觀感受程度，同時也蒐集受試者於研究期間每季一周主觀的整體情緒與作息紀錄。另一方面，透過腕動計(Actigraphy)記錄受試者於四季追蹤期間每日的客觀活動量強度(每分鐘一筆資料，每位受試者單次追蹤皆有10080筆數據)。我們首先將蒐集到的表現型數據進行傅立葉轉換(Fourier transform)，並使用熵數(sample entropy)、無母數轉換(Non-parameter transform)等統計方法，以了解不同診斷的受試者，其活動頻率和活動變異於相同季節中組間的差異。此外我們也比較疾病於四季之間組內變化的一致性，最後透過機器學習 – 隨機森林的方式，使用客觀的表徵數據做為診斷分類的依據，及是否正處於憂鬱狀態的特徵。研究結果用來進一步了解活動模式與情緒障礙間的相關性，並回答不同季節對於疾病與憂鬱狀態分類效果是否具有影響。 本研究發現在夏季時，重鬱症患者相較於健康對照組有較低的活動強度、較長的久坐時間以及較低的樣本熵，然而在秋季與冬季的夜間時段，與健康對照組相比，雙極性情感疾患患者有較低的活動強度、較長的久坐時間與較低的樣本熵。然而各組的活動量於春季期間則皆呈現相似的狀態。透過傅立葉轉換，我們發現在夏季與秋季時，重鬱症患者在第五及第九個每日最明顯的活動週期長度皆顯著低於健康照對組的數值。然而在季節間各組活動量變動的一致性當中，雙極性情感精神病和重鬱症患者在睡眠參數的一致性皆較健康對照組差。在GSS的部分，病例組在不同時間點評估季節變化對於活動量的影響具有較高的一致性。最後我們可以藉由活動以及睡眠的模式來有效區分不同疾病診斷及是否具有憂鬱狀態，且在四季中各有不同的最佳分類結果，在夏季時能較佳的區分重鬱症以及健康對照組(正確率: 0.81；敏感度: 0.86；特異度: 0.59)、在秋冬區分雙極性情感精神病以及健康對照組(正確率: 0.68 ~ 0.74；敏感度: 0.64 ~ 0.66；特異度: 0.47 ~ 0.66)、在春秋區分雙極性情感精神病以及重鬱症(正確率: 0.71 ~ 0.72；敏感度: 0.5 ~ 0.51；特異度: 0.74 ~ 0.75)皆有最好的效果，在區分憂鬱狀態時，在秋、冬、春季時利用BDI-II區分病例組與對照組是否為憂鬱狀態(正確率: 0.77 ~ 0.91；敏感度: 0.28 ~ 0.51；特異度: 0.85 ~ 0.91)、在春秋時利用BDI-II區分病例組是否為憂鬱症狀態(正確率: 0.7 ~ 0.73；敏感度: 0.56 ~ 0.67；特異度: 0.66 ~ 0.7)以及在四季利用睡眠日誌的憂鬱分數區分受試者是否為憂鬱狀態(正確率: 0.74 ~ 0.84；敏感度: 0.57 ~ 0.7；特異度: 0.71 ~ 0.91)也都有很好的分類效果。 最後總結，透過客觀的測量結果，我們發現客觀測量的活動量數據與診斷間在不同季節間沒有穩定的變動模式。也就是說，季節在探討診斷與活動模式、睡眠參數以及每日情緒變化之間的關係時可能扮演著修飾因子的角色。另一方面，在各個季節間使用主觀測量來評估季節性的活動或情緒狀態的結果可能會出現不一致。而藉由機器學習的方法，我們能夠對情緒障礙患者和健康對照者的診斷以及憂鬱狀態進行分類並達到可接受的準確度。未來若能夠同時使用客觀和主觀測量來記錄和評估情緒和活動模式可能可提供更全面的信息，以評估憂鬱症的嚴重程度和治療反應等重要變項。

Major depressive disorder and bipolar disorder are common and frequently relapsed mood disorders. Depressed is the common feature of mood disorders. According to previous studies, plenty of clinical symptoms are easily change with seasons. Among them, activity pattern and sleep status are the easiest observed phenotype data in our daily. Besides subjective description, if we have a completely objective phenotype dataset, it will help clinicians to diagnose patients more quickly, and evaluate the severity of episode, even in the future, it can detect the symptoms before the next episode onset. The present study mainly investigates the relationship between objective activity patterns and mood disorders and depressive symptoms between seasons, and we would use the phenotype data of activity and sleep that collected form actigraphy to classify different diagnoses between groups to reveal the relationships between clinical features and daily activity patterns in each season. The present study had 49 bipolar disorders and 52 major depressive disorders and recruited 41 healthy controls who didn’t have the medical history in psychiatry from communities to implement a one-week follow-up in each season. Among them, 53 participants have completed each follow-up in each season, and other participants collect follow-up data ranging from one to three seasons. We used Young Mania Rating Scale (YMRS), Hamilton Depression Rating Scale 17-item version (HAMD), and Beck Depression Inventory 21-item version (BDI-II) to evaluate the symptoms severity. This study also applied Seasonal Pattern Assessment Questionnaire(SPAQ) to assess the degree of subjective mood status in season changes, and simultaneously collect subjective mood status and daily routine for one week during follow-up period in each season. During seasonal follow-up, objective activity data were collected through actigraphy wore on participants' non-dominant hand for seven days and nights (each sample had 10080 data in each season). Firstly, we transformed the objective phenotype data with Fourier transform, and use sample entropy, non-parameter transforms and descriptive statistics to reveal the differences of activity frequency and fluctuation between groups in the same season. Besides, we also compare the consistency of fluctuation in each variable intra groups between four seasons. Lastly, through machine learning - random forest, using objective phenotype data as features to classify diagnoses and depressed status. With the results of the present study, we can further understand the relationship between objective activity patterns and mood disorders and explain whether seasons have an effect on the classification of diseases and depression. In the present study, patients with MDD had significantly lower activity count, longer sedentary time, and lower entropy than HC in summer, patients with BP had significantly lower activity count, longer sedentary time, and lower entropy than HC in the sleep time of autumn and winter. The activity patterns were similar in spring among all groups. Using Fourier transform, we found that MDD had significantly lower of the fifth and the ninth highest cycle per day than HC in summer and autumn. When it comes to consistency of activity fluctuation in different seasons, BP and MDD had lower consistency of sleep variables than HC. In the consistency of GSS score, case groups had higher consistency on evaluating the degree of seasonal effect on activity patterns in each season. Activity and sleep patterns can distinguish across diagnoses and depressed status. The better predicted performance was for distinguishing between MDD and HC in summer (accuracy : 0.81 ; sensitivity : 0.86 ; specificity : 0.59), between BP and HC in autumn and winter (accuracy ranged from 0.68 to 0.74 ; sensitivity ranged from 0.64 to 0.66 ; specificity ranged from 0.47 to 0.66), between BP and MDD in autumn and spring (accuracy ranged from 0.71 to 0.72 ; sensitivity ranged from 0.5 to 0.51 ; specificity ranged from 0.47 to 0.66). The better classified depressed status for distinguishing by BDI-II, based on cases and controls in autumn, winter, and spring (accuracy ranged from 0.77 to 0.91 ; sensitivity ranged from 0.28 to 0.51 ; specificity ranged from 0.85 to 0.91), based on case groups in the autumn and spring (accuracy ranged from 0.7 to 0.73; sensitivity ranged from 0.56 to 0.67 ; specificity ranged from 0.66 to 0.7). The better classified depressed status for distinguishing by sleep diary base on cases and controls or cases, in four seasons (accuracy ranged from 0.74 to 0.84; sensitivity ranged from 0.57 to 0.7; specificity ranged from 0.71 to 0.91). In conclusion, with objective measurements, we found that the relationship between objective activity data and disease groups were not stable across seasons. In other words, season may serve as a modifier for exploring the associations between diagnostic groups and activity patterns, sleep variables, and daily mood fluctuation. On the other hand, using subjective measurements to evaluate seasonal activity or mood status have inconsistent result between seasons. Using machine learning approach, we are able to classify diagnoses and depressive status among mood disorder patients and healthy controls with fair accuracy. In the future, if we can simultaneously use objective and subjective measurements to record and assess mood and activity pattern, we might can provide more comprehensive information to evaluate depressive severity, treatment response and other important index of mood disorders evaluations.