Integrating Neural Networks for Risk‐Adjustment Models

Abstract

This article demonstrates the possibility of an alternative approach for risk-adjustment models. In the proposed model the risk characteristics of the beneficiary's health within the same cohort classified by Self-Organizing Map network are highly homogeneous, whereas the numbers of individuals within each cohort remain sufficient to allow further investigation of the causal effect from clustered data. A comparison of different models by the 10-fold cross-validation reveals that the performance improvement in the proposed integration model is both significant and stable across the estimation and validation sampling. Copyright (c) The Journal of Risk and Insurance, 2008.

A review and comparison of medical expenditures models: two neural networks versus two‐part models

“…TNN possesses lower mean values of MSPE and MAD, a lower absolute value of MPB but higher mean values of PR and CC than those derived from TPM as well as from the SOM+LR model. This is consistent with the findings of Hsu, Lin, and Yang (2008) that integrating SOM and BPN performs better than TPM. However, we also find that although TNN has the smallest parameter bias, the smaller mean values of MSPE and MAD are associated with the greater standard deviation.…”

“…However, we also find that although TNN has the smallest parameter bias, the smaller mean values of MSPE and MAD are associated with the greater standard deviation. This also verifies the conjecture in Hsu, Lin, and Yang (2008) that the selection of BPN network architecture is important and undoubtedly influences model performance, because the functional relationships between input and output layers are described by the hidden nodes and hidden layers. Values of MPB and PR show that both TNN and SOM+LR yield slight underestimates in contrast to the overestimates from the TPM.…”

“…Brockett, Xia, and Derrig (1998) applied SOM to classify automobile injury claims and subsequently used BPN to examine the validity of the SOM classification approach. Hsu, Lin, and Yang (2008) predicted the annual individual medical expense by employing BPN with two different network structures in accordance with the data classification results of SOM. Although Hsu, Lin, and Yang (2008) showed a possible way for dealing with the above mentioned issues concerning the expense data and the two-stage decision process for medical demand, SOM serves as just a preparatory step for the BPN process that follows.…”

“…Hsu, Lin, and Yang (2008) predicted the annual individual medical expense by employing BPN with two different network structures in accordance with the data classification results of SOM. Although Hsu, Lin, and Yang (2008) showed a possible way for dealing with the above mentioned issues concerning the expense data and the two-stage decision process for medical demand, SOM serves as just a preparatory step for the BPN process that follows. Moreover, imposing each SOM risk cohort with one unique BPN may induce over-fitting since more unknown parameters need to be estimated.…”

Risk Assessment for Accounting Professional Liability Insurance

“…It was not only claimed that this assessment would affect insurance companies and positively reflect directly on their profitability, but it was also emphasized that risk assessment is crucial on the basis of policy (Smith et al, 2000: 532). Hsu, Lin and Yang (2008) developed a risk assessment model in health insurance using artificial neural networks. The assessment was performed using artificial neural networks with 7 risk factors.…”