Locally weighted factorization machine with fuzzy partition for elderly readmission prediction

The mitigation of preventable readmissions is the key to enhance the quality and efficiency of healthcare services in Hong Kong's aging society. A general predictor like factorization machine (FM) can efficiently capture non-linear interactions among patient characteristics for readmission prediction. However, FM cannot formulate local patient heterogeneity that commonly exists due to different demographics and hospitalization characteristics. This study proposes a locally weighted factorization machine (WFM) by developing a locally weighting scheme that assigns differentiated weights for training instances in accordance with their local similarities with high-readmission-risk ones. Io account for global heterogeneity among patient groups, we further extend WFM and propose a fuzzy partition enhanced WFM model (i.e., WFMFP), which (I) splits the dataset into multiple training subsets with fuzzy partition approach, (II) fits the WFM model on the basis of each training subset, and (III) combines the sub-prediction WFM models by using Takagi–Sugeno–Kang fuzzy weighting mechanism. We collected a territory-wide cohort with electronic health records in Hong Kong during 2008 to 2017. We demonstrated that the proposed WFMFP model outperforms several baselines, including standard FM, XGBoost, Lightgbm, Catboost, Gradient boosting machine, random forests, SVMs with different kernels, and multilayer perceptron. Comparative analysis results demonstrate the effectiveness of applying the locally weighting notion into FM to improve readmission prediction for clinical resource management.