A reliable time-series method for predicting arthritic disease outcomes: New step from regression toward a nonlinear
Hossein Bonakdari1, Jean-Pierre Pelletier1, Johanne Martel-Pelletier1
1Osteoarthritis Research Unit, University of Montreal Hospital Research Centre (CRCHUM), 900 rue Saint-Denis, R11.412, H2X 0A9, Montreal, Quebec, Canada.
Computer Methods and Programs in Biomedicine
|January 24, 2020
Summary
A new nonlinear method, Generalized Structure Group Method of Data Handling (GS-GMDH), offers a more accurate alternative to linear regression for interrupted time-series (ITS) analysis in public health. This AI-based approach simplifies data estimation and avoids the need to identify intervention change points.
Area of Science:
- Health policy analysis
- Biostatistics
- Artificial intelligence in medicine
Background:
- Interrupted time-series (ITS) analysis commonly uses linear regression to assess public health policy impacts.
- Traditional ITS methods require identifying specific change points and intervention lag times, which can be complex and time-consuming.
- Evaluating the impact of therapies, such as tumor necrosis factor inhibitors for rheumatoid arthritis, often relies on ITS analysis.
Purpose of the Study:
- To evaluate an artificial intelligence-based nonlinear approach, the Generalized Structure Group Method of Data Handling (GS-GMDH), for estimating ITS data.
- To determine if GS-GMDH can facilitate ITS data estimation and provide a computationally explicit equation.
- To compare the performance of GS-GMDH with traditional linear regression in a real-world health policy scenario.
Main Methods:
- Utilized a dataset evaluating the impact of NICE approval of tumor necrosis factor inhibitors on total hip replacement (THR) and total knee replacement (TKR) incidence in rheumatoid arthritis patients.
- Applied the Generalized Structure Group Method of Data Handling (GS-GMDH), a nonlinear AI model, for predicting THR and TKR incidence.
- Compared GS-GMDH predictions against measured data and linear regression models.
Main Results:
- GS-GMDH models demonstrated superior prediction accuracy for both THR and TKR compared to linear regression, with low mean absolute relative errors (0.10 for THR, 0.09 for TKR) and high correlation coefficients (0.98 for THR, 0.78 for TKR).
- GS-GMDH provided significantly different post-intervention incidence estimates compared to linear regression (e.g., THR: 6.4/1000 PYs vs. 4.12/1000 PYs; TKR: 12.47/1000 PYs vs. 9.05/1000 PYs).
- A key advantage of GS-GMDH is its ability to continually simulate ITS without requiring pre-identification of change points or intervention lag times.
Conclusions:
- The nonlinear GS-GMDH method is a more accurate and efficient alternative to traditional regression-based methods for processing interrupted time-series data in medical research.
- GS-GMDH offers improved prediction accuracy and reduces the need for time-consuming experimental measurements.
- This AI-driven nonlinear approach overcomes a significant challenge in ITS modeling by eliminating the necessity to identify change points and intervention lag times.
