Mortality risk estimation for autistic older adults: comparing novel machine-learning derived weights versus standard...

Area of Science:

• Gerontology
• Autism Spectrum Disorder Research
• Health Services Research

Background:

• The Charlson Comorbidity Index (CCI) is widely used to predict mortality risk.
• Its established weights may not accurately reflect mortality predictors in autistic older adults.
• There is a need to evaluate and potentially adapt the CCI for specific populations.

Purpose of the Study:

• To compare established CCI weights against autism-specific weights for predicting mortality in older adults with autism.
• To assess the predictive performance of both weighting systems for short-term (30-day) and long-term (1-year) mortality.

Main Methods:

• Utilized Medicare inpatient healthcare claims data from 2829 autistic older adults (aged 65+).
• Employed stochastic hill climbing, a machine learning technique, to derive autism-specific weights for 12 CCI conditions.
• Compared the predictive accuracy (Area Under the Curve - AUC) of established versus autism-specific weights.

Main Results:

• Both established and autism-specific CCI weights demonstrated poor predictive ability for 30-day mortality (AUC: 0.68 and 0.67, respectively).
• Predictive performance for 1-year mortality was also poor for both weighting systems (AUC: 0.67 for both).
• The AUC values indicated no significant difference in predictive accuracy between the two weighting approaches.

Conclusions:

• Neither established nor autism-specific CCI weights accurately predict mortality in autistic older adults.
• Adjusting CCI weights alone is insufficient; additional, non-CCI health conditions may be necessary for improved risk prediction.
• Further research is warranted to develop a dedicated autism-specific mortality risk index.