Abstract: Business analytics has become an increasingly important priority for organizations today as they strive to achieve greater competitiveness. As organizations adopt business practices that rely on complex, large-scale data, new challenges also emerge. A common situation in business analytics is concerned with appropriate and adequate methods for dealing with unlabeled data in classification. This study examines the effectiveness of a semi-supervised learning approach to classify unlabeled data to improve classification accuracy rates. The context for our study is healthcare. The healthcare costs in the U.S. have risen at an alarming rate over the last two decades. One of the causes for the rising costs could be attributed to medical bad debt, i.e., debt that is not recovered by healthcare institutions. A major obstacle to debt classification, hence better debt recovery, is the presence of unlabeled cases, a situation not uncommon in many other business contexts. There is surprisingly very little research that explores the performance of computational intelligence and soft computing methods in improving bad debt recovery in the healthcare industry. Using a real data set from a healthcare organization, we address this important research gap by examining the performance of an adaptive neuro-fuzzy inference system (ANFIS) with semi-supervised learning (SSL) in improving debt recovery rate. In particular, this study explores the role of ANFIS in conjunction with SSL in classifying unknown cases (those that were not pursued for debt collection) as either a good case (recoverable) or a bad case (unrecoverable). Healthcare institutions can then pursue these potentially good cases and improve their debt recovery rates. Test results show that ANFIS with SSL is a viable method. Our models generated better classification accuracy rates than those in prior studies. These results and their analysis show the potential of ANFIS with SSL models in classifying unknown cases, which are a potential source of revenue recovery for health care organizations. The significance of this research extends to all types of organizations that face an increasingly urgent need to adopt reliable practices for business analytics.
Introduction
In the U.S. healthcare sector, medical bad debt—unpaid bills that hospitals eventually stop trying to collect—is a massive drain on resources. A major hurdle for hospitals is the "unlabeled data" problem. While they have records of debts that were successfully recovered ("good") and those that were pursued but failed ("bad"), they also have a massive category of "unknown" cases: debts that were never pursued, often because they were too small or the patient’s financial profile was unclear.
Traditional classifiers struggle here because they require labeled data to learn. Furthermore, the data is often "imbalanced," meaning there are far fewer successful recoveries than failures. This study examines the effectiveness of a semi-supervised learning approach to classify unlabeled data to improve classification accuracy rates, specifically targeting the recovery of medical bad debt.
Methodology
The researchers proposed a novel ANFIS-based self-training algorithm. This approach integrates two powerful computational paradigms:
- ANFIS (Adaptive Neuro-Fuzzy Inference System): This system combines the transparency of Fuzzy Logic (using "if-then" rules to handle uncertainty) with the computational power of Neural Networks (which "learn" the best parameters from data).
- Semi-Supervised Learning (SSL): Instead of ignoring the "unknown" cases, the model uses a "self-training" wrapper. It starts by training on a balanced set of known "good" and "bad" cases. It then predicts the labels for the "unknown" data. The most confident predictions are iteratively added back into the training set, allowing the model to "teach itself" from the unlabeled pool.
As illustrated in Figure 1, the ANFIS architecture consists of five distinct layers. Layer 1 maps inputs (such as patient age or claim amount) to fuzzy sets using membership functions. Layer 2 performs the "AND" operation for the fuzzy rules, while Layer 3 normalizes the firing strengths. Layer 4 calculates the consequent part of the rules, and Layer 5 computes the final output as a summation of all incoming signals, leading to a classification decision.
Experimental Results
The study utilized a real-world dataset from a healthcare organization containing 6,117 cases with a total outstanding balance of over $2.3 million. The model relied on four primary inputs: Patient Age (PA), Patient Gender (PG), Injury Diagnosis Code (IDC), and the Dollar Amount of the Claim (DAC).
The ANFIS with SSL approach significantly outperformed previous benchmarks:
- Accuracy: The hybrid model achieved overall correct classification rates between 74.8% and 76.7%, depending on the specific fuzzy membership function used.
- Comparison: This was a marked improvement over ANFIS alone (which hovered around 61-63% in previous scenarios) and outperformed traditional Neural Networks (72.3%) and Logistic Regression (75.0%) from prior studies.
- Recovery Potential: After applying SSL, the model reclassified 1,208 of the "unknown" cases as "good," providing a clear roadmap for the hospital to target its collection efforts more effectively. These cases represented a potential $548,373 in additional recoverable revenue.
Conclusion
The results demonstrate the potential of ANFIS with SSL models in classifying unknown cases, which are a potential source of revenue recovery for healthcare organizations. The use of fuzzy logic allowed the model to handle the inherent uncertainty in predicting human payment behavior, while the semi-supervised learning component successfully leveraged unlabeled data to improve performance.
Future research could incorporate more granular data—like socio-economic status or more detailed medical histories—to further refine accuracy. The significance of this research extends beyond healthcare to all types of organizations that face an increasingly urgent need to adopt reliable practices for business analytics in the presence of incomplete or unlabeled data.
Summary updated