A comparison of machine learning classifiers for pediatric epilepsy using resting-state functional MRI latency data.

Epilepsy affects 1 in 150 children under the age of 10 and is the most common chronic pediatric neurological condition; poor seizure control can irreversibly disrupt normal brain development. The present study compared the ability of different machine learning algorithms trained with resting-state functional MRI (rfMRI) latency data to detect epilepsy. Preoperative rfMRI and anatomical MRI scans were obtained for 63 patients with epilepsy and 259 healthy controls. The normal distribution of latency z-scores from the epilepsy and healthy control cohorts were analyzed for overlap in 36 seed regions. In these seed regions, overlap between the study cohorts ranged from 0.44-0.58. Machine learning features were extracted from latency z-score maps using principal component analysis. Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), and Random Forest algorithms were trained with these features. Area under the receiver operating characteristics curve (AUC), accuracy, sensitivity, specificity and F1-scores were used to evaluate model performance. The XGBoost model outperformed all other models with a test AUC of 0.79, accuracy of 74%, specificity of 73%, and a sensitivity of 77%. The Random Forest model performed comparably to XGBoost across multiple metrics, but it had a test sensitivity of 31%. The SVM model did not perform >70% in any of the test metrics. The XGBoost model had the highest sensitivity and accuracy for the detection of epilepsy. Development of machine learning algorithms trained with rfMRI latency data could provide an adjunctive method for the diagnosis and evaluation of epilepsy with the goal of enabling timely and appropriate care for patients.

Convolutional Neural Networks for Pediatric Refractory Epilepsy Classification Using Resting-State Functional Magnetic Resonance Imaging.

OBJECTIVE: This study aims to evaluate the performance of convolutional neural networks (CNNs) trained with resting-state functional magnetic resonance imaging (rfMRI) latency data in the classification of patients with pediatric epilepsy from healthy controls. METHODS: Preoperative rfMRI and anatomic magnetic resonance imaging scans were obtained from 63 pediatric patients with refractory epilepsy and 259 pediatric healthy controls. Latency maps of the temporal difference between rfMRI and the global mean signal were calculated using voxel-wise cross-covariance. Healthy control and epilepsy latency z score maps were pseudorandomized and partitioned into training data (60%), validation data (20%), and test data (20%). Healthy control individuals and patients with epilepsy were labeled as negative and positive, respectively. CNN models were then trained with the designated training data. Model hyperparameters were evaluated with a grid-search method. The model with the highest sensitivity was evaluated using unseen test data. Accuracy, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve were used to evaluate the ability of the model to classify epilepsy in the test data set. RESULTS: The model with the highest validation sensitivity correctly classified 74% of unseen test patients with 85% sensitivity, 71% specificity, F1 score of 0.56, and an area under the receiver operating characteristic curve of 0.86. CONCLUSIONS: Using rfMRI latency data, we trained a CNN model to classify patients with pediatric epilepsy from healthy controls with good performance. CNN could serve as an adjunct in the diagnosis of pediatric epilepsy. Identification of pediatric epilepsy earlier in the disease course could decrease time to referral to specialized epilepsy centers and thus improve prognosis in this population.

Interobserver Agreement Between Pathologists Assessing Tumor-Infiltrating Lymphocytes (TILs) in Breast Cancer Using Methodology Proposed by the International TILs Working Group.

BACKGROUND: The presence of tumor-infiltrating lymphocytes (TILs) in breast tumors is prognostic and predictive, suggesting that TILs may be an important biomarker. Recently, an international TILs working group formulated consensus recommendations for TIL evaluation. The current study was performed to determine interobserver agreement using that methodology. METHODS: Tumor-infiltrating lymphocytes were assessed on a single hematoxylin and eosin (H&E)-stained slide obtained from the core biopsy of 75 triple-negative breast cancers. Four pathologists independently reviewed each slide and evaluated stromal TILs (sTILs) and intratumoral TIL (iTILs). The kappa statistic was used to estimate interobserver agreement for identification of sTILs, and the intraclass correlation coefficient (ICC) was used to estimate the agreement among observers for iTILs. Cases with poor agreement were reviewed to identify pathologic factors that may contribute to the lack of agreement. RESULTS: The kappa statistic for sTIL evaluation was 0.57 (standard error, 0.04). For iTILs, the ICC calculated to determine internal consistency within raters was 0.65 (95 % confidence interval [CI] 0.56-0.74; p < 0.0001), and the ICC calculated to determine agreement among raters was 0.62 (95 % CI 0.50-0.72; p < 0.0001). In 10 cases (13 %), there was not agreement between three of four pathologists. The pathologic features contributing to difficulty in TIL enumeration included marked individual tumor cell necrosis or apoptosis, the presence of reactive plasma cells mimicking tumor cells, plasmatoid tumor cells, and accurate quantification of TILs in specimens with focal areas of heavy immune infiltrate. CONCLUSION: Acceptable agreement in TIL enumeration was observed, suggesting that the proposed methodology can be used to facilitate the use of TILs as a biomarker in research and clinical trial settings.

Locoregional Control According to Breast Cancer Subtype and Response to Neoadjuvant Chemotherapy in Breast Cancer Patients Undergoing Breast-conserving Therapy.

BACKGROUND: Our group previously published data showing that patients could be stratified by constructed molecular subtype with respect to locoregional recurrence (LRR)-free survival after neoadjuvant chemotherapy and breast-conserving therapy (BCT). That study predated use of trastuzumab for human epidermal growth factor receptor 2 (HER2)-positive patients. The current study was undertaken to determine the impact of subtype and response to therapy in a contemporary cohort. METHODS: Clinicopathologic data from 751 breast cancer patients who received neoadjuvant chemotherapy (with trastuzumab if HER2(+)) and BCT from 2005 to 2012 were identified. Hormone receptor (HR) and HER2 status were used to construct molecular subtypes: HR(+)/HER2(-) (n = 369), HR(+)/HER2(+) (n = 105), HR(-)/HER2(+) (n = 58), and HR(-)/HER2(-) (n = 219). Actuarial rates of LRR were determined by the Kaplan-Meier method and compared by the log-rank test. Multivariate analysis was performed to determine factors associated with LRR. RESULTS: The pathologic complete response (pCR) rates by subtype were as follows: 16.5% (HR(+)/HER2(-)), 45.7% (HR(+)/HER2(+)), 72.4% (HR(-)/HER2(+)), and 42.0% (HR(-)/HER2(-)) (P < 0.001). Median follow-up was 4.6 years. The 5-year LRR-free survival rate for all patients was 95.4%. Five-year LRR-free survival rates by subtype were 97.2 % (HR(+)/HER2(-)), 96.1% (HR(+)/HER2(+)), 94.4% (HR(-)/HER2(+)), and 93.4% (HR(-)/HER2(-)) (P = 0.44). For patients with HR(-)/HER2(+) disease, the LRR-free survival rates were 97.4 and 86.7% for those who did and those who did not experience pCR, respectively. For patients with HR(-)/HER2(-) disease, the LRR-free survival rates were 98.6% (pCR) versus 89.9% (no pCR). On multivariate analysis, the HR(-)/HER2(-) subtype, clinical stage III disease, and failure to experience a pCR were associated with LRR. CONCLUSIONS: Patients undergoing BCT after neoadjuvant chemotherapy have excellent rates of 5-year LRR-free survival that are affected by molecular subtype and by response to neoadjuvant chemotherapy.