| 摘要: | 背景
中風是全世界長期殘疾的主要原因之一，認知障礙是中風的常見後果。它可以影響認知功能的各個領域，例如注意力、記憶、語言和執行功能。中風後認知障礙（PSCI）被定義為中風後認知能力下降，影響患者進行日常生活活動的能力。 PSCI 對中風倖存者及其護理人員的生活質量具有重大影響。 PSCI 的患病率和風險 研究報告稱，PSCI 的患病率範圍很廣，從 25% 到 81%，具體取決於年齡、中風嚴重程度和合併症等各種因素。研究發現，中風後癡呆的風險隨著時間的推移而增加，中風後癡呆的發病率逐年增加。然而，中風後患癡呆症的相對風險隨著時間的推移逐漸降低。鑑於認知障礙對中風倖存者的重大影響，儘早識別中風後有發生認知障礙風險的患者對於提供適當的護理和管理至關重要。這項研究旨在使用機器學習方法來預測中風診斷一年後認知障礙的發展。
方法
該研究使用 2004 年 1 月至 2017 年 9 月的數據集在 TMU-CRD 上進行。該研究的納入和排除標準是根據 ICD9 和 ICD10 代碼確定的。該研究納入了有中風、失眠和認知障礙病史以及其他與疾病相關的代碼的患者。然而，患有精神疾病、睡眠呼吸暫停、創傷性腦損傷、癌症和帕金森病的患者被排除在結果之外。該研究的結果測量由多種類型組成，包括輕度認知障礙、老年癡呆（無並發症）、伴有妄想或抑鬱特徵的老年癡呆、伴有譫妄的老年癡呆、分類為其他疾病的癡呆、阿爾茨海默病、額顳葉癡呆和老年癡呆大腦退化。為了處理、訓練、測試和驗證數據，使用了 MySQL 和 PyCaret 庫。最後，為了確認結果，使用了 OHDSI 的其他框架和數據集格式、R 包，即 OHDSI 框架中的 PatientLevelPrediction 和 DeepPatientLevelPrediction。
結果
獲得的分數如下：準確度分數為0.95，AUC分數為0.85，精確度分數為0.55，召回分數為0.10，F1分數為0.16。 OHDSI框架和OMOP CDM數據集的確認結果為準確度得分為0.96，AUC得分為0.81，精確度得分為0.17，召回得分為0.23，F1得分為0.19。值得注意的是，在當前的分析中，僅使用診斷代碼、性別/年齡和藥物作為預測卒中診斷後一年內認知障礙發展的特徵。研究結果表明，光梯度增強機（LightGBM）可以有效預測中風患者早期認知障礙的發展。然而，可能需要對附加功能進行進一步研究以提高模型的性能。該研究的結果可能對中風患者認知障礙和癡呆的早期識別和預防具有重要意義，從而改善患者的治療結果和提高生活質量。
結論
現有模型能夠預測患者在接下來的一年中發生認知障礙的情況，但這種情況發生的概率仍然很低。該模型依靠性別、年齡、診斷和藥物作為進行預測的關鍵特徵。然而，有機會通過實施額外的功能工程來增強其性能。為了實現這一目標，我們將納入其他可能影響認知功能的變量，例如程序、測量和實驗室測試。這些附加功能將集成到模型中以提高其預測能力。通過這樣做，有望提高真陽性率，同時降低假陰性率，從而提供更準確的認知障礙風險預測。此外，通過採用 OHDSI 框架和工具，全球多中心以及多國協作是廣泛開放的。
Background
Stroke is one of the leading causes of long-term disability worldwide, and cognitive impairment is a common consequence of stroke. It can affect various domains of cognitive function such as attention, memory, language, and executive function. Post-stroke cognitive impairment (PSCI) is defined as cognitive decline following a stroke that affects the patient's ability to perform activities of daily living. PSCI has a significant impact on the quality of life of stroke survivors and their caregivers. Prevalence and Risk of PSCI Studies have reported a wide prevalence range of PSCI, from 25% to 81%, depending on various factors such as age, stroke severity, and comorbidities. Studies have found that the risk of dementia after a stroke increases over time, with the incidence rate of post-stroke dementia increasing yearly. However, the relative risk of developing dementia after a stroke gradually decreases over time. Given the significant impact of cognitive impairment on stroke survivors, identifying patients who are at risk of developing cognitive impairment as early as possible after a stroke is crucial for providing appropriate care and management. This study aims to use a machine learning approach to predict cognitive impairment development one year after stroke diagnoses.
Methods
Using a dataset from January 2004 to September 2017, the study was performed on TMU-CRD. The inclusion and exclusion criteria for the study were determined based on the ICD9 and ICD10 codes. Patients who had a history of stroke, insomnia, and cognitive impairment, as well as other codes related to the diseases, were included in the study. However, patients with psychiatric disorders, sleep apnea, traumatic brain injury, cancer, and Parkinson's disease were excluded from the outcome. The outcome measures for the study were consists of several types, including mild cognitive impairment, Senile dementia (uncomplicated), Senile dementia with delusional or depressive features, Senile dementia with delirium, Dementia in conditions classified elsewhere, Alzheimer's disease, Frontotemporal dementia, and Senile degeneration of the brain. To process, train, test and validate the data, MySQL and PyCaret library were used. Finally, to confirm the result, other framework and dataset format from OHDSI, R packages namely PatientLevelPrediction and DeepPatientLevelPrediction from the OHDSI frameworks were utilized.
Result
The scores obtained were as follows: an accuracy score of 0.95, an AUC score of 0.85, a precision score of 0.55, a recall score of 0.10, and an F1 score of 0.16. Confirmation result from OHDSI framework and OMOP CDM dataset were accuracy score of 0.96, AUC score of 0.81, precision score of 0.17, recall score of 0.23 and F1 score of 0.19. It should be noted that in the current analysis, only diagnosis codes, gender/age, and medication were used as the features to predict cognitive impairment development within one year after stroke diagnosis. The study's results suggest that the light gradient boosting machine (LightGBM) can effectively predict early cognitive impairment development among stroke patients. However, further investigation with additional features may be necessary to improve the model's performance. The study's findings may have significant implications for the early identification and prevention of cognitive impairment and dementia in stroke patients, leading to improved patient outcomes and better quality of life.
Conclusion
The existing model has the capability to predict the occurrence of cognitive impairment in patients over the following year, but the probability of this happening remains low. The model relies on gender, age, and diagnosis and medication as the key features to make the prediction. However, there is opportunity to enhance its performance by implementing additional features engineering. To achieve this objective, we will incorporate other variables that can affect cognitive function, such as procedures, measurements and laboratory tests. These additional features will be integrated into the model to improve its predictive capabilities. By doing so, it is expected to enhance the true positive rate while reducing the false negative rate, which would provide a more accurate prediction of cognitive impairment risk. Furthermore, by employing the OHDSI framework and tooling, the global multicenter, as well as multi-country collaboration is widely open. |
