Research on the US urban system has shown that metropolitan regions with more local and non-local network ties outperform cities where economic agents are isolated. Yet, little attention is given to the character of the local knowledge base and how that influences network structure. We show that co-inventor networks differ between cities that produce specialized and diversified knowledge. Models of tie-formation show inventors in specialized cities value spatial proximity less and cognitive proximity more than inventors in diversified cities as they partner with nonlocal inventors. These findings suggest that the influence of social networks on knowledge production is conditioned by the architecture of the local knowledge base.

INTRODUCTION Reliable preoperative identification of patients at high risk for early postoperative complications occurring within 24 h (EPC) of intracranial tumor surgery can improve patient safety and postoperative management. Statistical analysis using machine learning algorithms may generate models that predict EPC better than conventional statistical methods. OBJECTIVE To train such a model and to assess its predictive ability. METHODS This cohort study included patients from an ongoing prospective patient registry at a single tertiary care center with an intracranial tumor that underwent elective neurosurgery between June 2015 and May 2017. EPC were categorized based on the Clavien-Dindo classification score. Conventional statistical methods and different machine learning algorithms were used to predict EPC using preoperatively available patient, clinical, and surgery-related variables. The performance of each model was derived from examining classification performance metrics on an out-of-sample test dataset. RESULTS EPC occurred in 174 (26%) of 668 patients included in the analysis. Gradient boosting machine learning algorithms provided the model best predicting the probability of an EPC. The model scored an accuracy of 0.70 (confidence interval [CI] 0.59-0.79) with an area under the curve (AUC) of 0.73 and a sensitivity and specificity of 0.80 (CI 0.58-0.91) and 0.67 (CI 0.53-0.77) on the test set. The conventional statistical model showed inferior predictive power (test set: accuracy: 0.59 (CI 0.47-0.71); AUC: 0.64; sensitivity: 0.76 (CI 0.64-0.85); specificity: 0.53 (CI 0.41-0.64)). CONCLUSION Using gradient boosting machine learning algorithms, it was possible to create a prediction model superior to conventional statistical methods. While conventional statistical methods favor patients’ characteristics, we found the pathology and surgery-related (histology, anatomical localization, surgical access) variables to be better predictors of EPC.