ABSTRACT
PURPOSE: This study aimed to subtype multiple sclerosis (MS) patients using unsupervised machine learning on white matter (WM) fiber tracts and investigate the implications for cognitive function and disability outcomes. MATERIALS AND METHODS: We utilized the automated fiber quantification (AFQ) method to extract 18 WM fiber tracts from the imaging data of 103 MS patients in total. Unsupervised machine learning techniques were applied to conduct cluster analysis and identify distinct subtypes. Clinical and diffusion tensor imaging (DTI) metrics were compared among the subtypes, and survival analysis was conducted to examine disability progression and cognitive impairment. RESULTS: The clustering analysis revealed three distinct subtypes with variations in fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD). Significant differences were observed in clinical and DTI metrics among the subtypes. Subtype 3 showed the fastest disability progression and cognitive decline, while Subtype 2 exhibited a slower rate, and Subtype 1 fell in between. CONCLUSIONS: Subtyping MS based on WM fiber tracts using unsupervised machine learning identified distinct subtypes with significant cognitive and disability differences. WM abnormalities may serve as biomarkers for predicting disease outcomes, enabling personalized treatment strategies and prognostic predictions for MS patients.
ABSTRACT
PURPOSE: The aim of this meta-analysis was to determine the diagnostic accuracy of machine learning (ML) models with MRI in predicting pathological response to neoadjuvant chemotherapy in patients with breast cancer. Furthermore, we compared the pathologic complete response (pCR) prediction performance of ML + radiomics with that of a deep learning (DL) algorithm. METHODS: A search for relevant studies published until December 20, 2021 was conducted in MEDLINE and EMBASE databases. The quality of the studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies -2 criteria. The I2 value assessed the heterogeneity of the included studies as well as the decision to adopt a random effects model. The area under the receiver operating characteristic curves (AUC) was pooled to quantify the predictive accuracy. Subgroup analysis, meta-regression analysis, and sensitivity analysis were performed to detect potential sources of study heterogeneity. A funnel plot was used to investigate publication bias. The PROSPERO ID of our study was CRD42022284071. RESULT: Seventeen eligible studies encompassing 3392 patients were evaluated in the analysis. ML + MRI showed high accuracy (AUC = 0.87, 95% CI = 0.84-0.91) in predicting response to neoadjuvant therapy. In subgroup analysis, the AUC of the DL subgroup (AUC = 0.92, 95% CI = 0.88-0.97) was higher than that of the ML + radiomics subgroup (AUC = 0.85, 95% CI = 0.82-0.90) (P = 0.030). In the ML + radiomics subgroup, the studies using MRI combined with other parameters (clinical or histopathologic information; AUC = 0.90, 95% CI = 0.85-0.96) reported better performance than studies using only MRI parameters (AUC = 0.82, 95% CI = 0.78-0.86) (P = 0.009). CONCLUSIONS: ML applied to MRI enabled moderate accuracy in predicting pathological response to neoadjuvant therapy in patients with breast cancer. Furthermore, the meta-analysis showed that DL had higher predictive accuracy than ML + radiomics.
