Deriving injury risk curves using survival analysis from biomechanical experiments

Abstract

Injury risk curves from biomechanical experimental data analysis are used in automotive studies to improve crashworthiness and advance occupant safety. Metrics such as acceleration and deflection coupled with outcomes such as fractures and anatomical disruptions from impact tests are used in simple binary regression models. As an improvement, the International Standards Organization suggested a different approach. It was based on survival analysis. While probability curves for side-impact-induced thorax and abdominal injuries and frontal impact-induced foot-ankle-leg injuries are developed using this approach, deficiencies are apparent. The objective of this study is to present an improved, robust and generalizable methodology in an attempt to resolve these issues. It includes: (a) statistical identification of the most appropriate independent variable (metric) from a pool of candidate metrics, measured and or derived during experimentation and analysis processes, based on the highest area under the receiver operator curve, (b) quantitative determination of the most optimal probability distribution based on the lowest Akaike information criterion, (c) supplementing the qualitative/visual inspection method for comparing the selected distribution with a non-parametric distribution with objective measures, (d) identification of overly influential observations using different methods, and (e) estimation of confidence intervals using techniques more appropriate to the underlying survival statistical model. These clear and quantified details can be easily implemented with commercial/open source packages. They can be used in retrospective analysis and prospective design of experiments, and in applications to different loading scenarios such as underbody blast events. The feasibility of the methodology is demonstrated using post mortem human subject experiments and 24 metrics associated with thoracic/abdominal injuries in side-impacts.

Bayesian Stacked Parametric Survival with Frailty Components and Interval‐Censored Failure Times: An Application to Food Allergy Risk

“…When estimating risk to populations exposed to environmental hazards, interval‐censored failure data are common. Such data can be found in many areas of research including animal carcinogenicity experiments, demographical, epidemiological, financial, medical, sociological, machine reliability, and injury risk studies (Collett, 2015; Crowder, Crouse, Heppell, & Martin, 1994; Klein, 1992; Klein, Van Houwelingen, Ibrahim, & Scheike, 2016; Ozturk et al., 2018; Sun et al., 2006; Yoganandan et al., 2016). Typical examples of interval‐censored data are found in medical or health studies with periodic follow‐up protocols, with examples including time‐to‐tumor or time‐to‐infection studies (e.g., appearance of lung cancer tumors, breast cancer studies with four‐ to six‐month follow‐up times, timing of HIV infection in AIDS cohort studies with follow‐up every six months) (Sun et al., 2006).…”

Rib Cortical Bone Fracture Risk as a Function of Age and Rib Strain: Updated Injury Prediction Using Finite Element Human Body Models

“…AIC was used to select the log-normal distribution. It has been debated whether AIC is suitable for choosing the distribution, or if one should default to a Weibull distribution, or if the Area under the Receiver Operator Curve, indicating how good injury and non-injury data are classified, is a better metric ( Yoganandan et al, 2016 , 2017 ; McMurry and Poplin, 2017 ). For the developed risk curves the Weibull distribution performed worst in terms of AIC, but with an AIC delta of less than ten compared to the other distributions.…”

“…Using the newly developed risk function it is predicted at a Delta-v of 52 km/h, an underestimation of 9 km/h. For a 70-year-old occupant, the corresponding underestimations of the Delta-v for the 50% Curve, indicating how good injury and non-injury data are classified, is a better metric (Yoganandan et al, 2016(Yoganandan et al, , 2017McMurry and Poplin, 2017). For the developed risk curves the Weibull distribution performed worst in terms of AIC, but with an AIC delta of less than ten compared to the other distributions.…”

Biomechanical Responses and Injury Characteristics of Knee Joints under Longitudinal Impacts of Different Velocities

“…Impacting injury to the knee joint is most commonly seen in traffic injuries [ 27 ] or sports injuries [ 28 ]. Most studies have been conducted under knee bending conditions [ 29 ], and there are relatively few studies on the mechanisms and characteristics of impact injury, especially for knee injuries induced by longitudinal impacts.…”