Abstract：
BACKGROUND: Differentiating Progressive Supranuclear Palsy (PSP) from Parkinson\'s Disease (PD) may be clinically challenging. In this study, we explored the performance of machine learning models based on MR imaging and blood molecular biomarkers in distinguishing between these two neurodegenerative diseases.
METHODS: Twenty-eight PSP patients, 46 PD patients and 60 control subjects (HC) were consecutively enrolled in the study. Serum concentration of neurofilament light chain protein (Nf-L) was assessed by single molecule array (SIMOA), while an automatic segmentation algorithm was employed for T1-weighted measurements of third ventricle width/intracranial diameter ratio (3rdV/ID). Machine learning (ML) models with Logistic Regression (LR), Random Forest (RF), and XGBoost algorithms based on 3rdV/ID and serum Nf-L levels were tested in distinguishing among PSP, PD and HC.
RESULTS: PSP patients showed higher serum Nf-L levels and larger 3rdV/ID ratio in comparison with both PD and HC groups (p < 0.005). All ML algorithms (LR, RF and XGBoost) showed that the combination of MRI and blood biomarkers had excellent classification performances in differentiating PSP from PD (AUC ≥0.92), outperforming each biomarker used alone (AUC: 0.85-0.90). Among the different algorithms, XGBoost was slightly more powerful than LR and RF in distinguishing PSP from PD patients, reaching AUC of 0.94 ± 0.04.
CONCLUSIONS: Our findings highlight the usefulness of combining blood and simple linear MRI biomarkers to accurately distinguish between PSP and PD patients. This multimodal approach may play a pivotal role in patient management and clinical decision-making, paving the way for more effective and timely interventions in these neurodegenerative diseases.
METHODS: Twenty-eight PSP patients, 46 PD patients and 60 control subjects (HC) were consecutively enrolled in the study. Serum concentration of neurofilament light chain protein (Nf-L) was assessed by single molecule array (SIMOA), while an automatic segmentation algorithm was employed for T1-weighted measurements of third ventricle width/intracranial diameter ratio (3rdV/ID). Machine learning (ML) models with Logistic Regression (LR), Random Forest (RF), and XGBoost algorithms based on 3rdV/ID and serum Nf-L levels were tested in distinguishing among PSP, PD and HC.
RESULTS: PSP patients showed higher serum Nf-L levels and larger 3rdV/ID ratio in comparison with both PD and HC groups (p < 0.005). All ML algorithms (LR, RF and XGBoost) showed that the combination of MRI and blood biomarkers had excellent classification performances in differentiating PSP from PD (AUC ≥0.92), outperforming each biomarker used alone (AUC: 0.85-0.90). Among the different algorithms, XGBoost was slightly more powerful than LR and RF in distinguishing PSP from PD patients, reaching AUC of 0.94 ± 0.04.
CONCLUSIONS: Our findings highlight the usefulness of combining blood and simple linear MRI biomarkers to accurately distinguish between PSP and PD patients. This multimodal approach may play a pivotal role in patient management and clinical decision-making, paving the way for more effective and timely interventions in these neurodegenerative diseases.
摘要：
背景：区分进行性核上性麻痹（PSP）和帕金森病（PD）可能在临床上具有挑战性。在这项研究中,我们探索了基于MR成像和血液分子生物标志物的机器学习模型在区分这两种神经退行性疾病方面的表现.
方法：28例PSP患者，46名PD患者和60名对照受试者（HC）连续纳入研究。通过单分子阵列（SIMOA）评估神经丝轻链蛋白（Nf-L）的血清浓度，而第三脑室宽度/颅内直径比（3rdV/ID）的T1加权测量采用自动分割算法。具有Logistic回归(LR)的机器学习(ML)模型，随机森林(RF),基于3rdV/ID和血清NF-L水平的XGBoost算法在区分PSP中进行了测试，PD和HC。
结果：与PD和HC组相比，PSP患者的血清Nf-L水平更高，3rdV/ID比率更高（p<0.005）。所有ML算法(LR，RF和XGBoost）表明，MRI和血液生物标志物的组合在区分PSP与PD（AUC≥0.92）方面具有出色的分类性能，优于单独使用的每种生物标志物(AUC:0.85-0.90)。在不同的算法中，在区分PSP和PD患者方面，XGBoost比LR和RF功能稍强。AUC达到0.94±0.04。
结论：我们的研究结果强调了将血液和简单的线性MRI生物标志物结合在一起以准确区分PSP和PD患者的有用性。这种多模式方法可能在患者管理和临床决策中发挥关键作用。为更有效和及时地干预这些神经退行性疾病铺平道路。
方法：28例PSP患者，46名PD患者和60名对照受试者（HC）连续纳入研究。通过单分子阵列（SIMOA）评估神经丝轻链蛋白（Nf-L）的血清浓度，而第三脑室宽度/颅内直径比（3rdV/ID）的T1加权测量采用自动分割算法。具有Logistic回归(LR)的机器学习(ML)模型，随机森林(RF),基于3rdV/ID和血清NF-L水平的XGBoost算法在区分PSP中进行了测试，PD和HC。
结果：与PD和HC组相比，PSP患者的血清Nf-L水平更高，3rdV/ID比率更高（p<0.005）。所有ML算法(LR，RF和XGBoost）表明，MRI和血液生物标志物的组合在区分PSP与PD（AUC≥0.92）方面具有出色的分类性能，优于单独使用的每种生物标志物(AUC:0.85-0.90)。在不同的算法中，在区分PSP和PD患者方面，XGBoost比LR和RF功能稍强。AUC达到0.94±0.04。
结论：我们的研究结果强调了将血液和简单的线性MRI生物标志物结合在一起以准确区分PSP和PD患者的有用性。这种多模式方法可能在患者管理和临床决策中发挥关键作用。为更有效和及时地干预这些神经退行性疾病铺平道路。
