AutoScore-Survival: Developing interpretable machine learning-based time-to-event scores with right-censored survival...

Area of Science:

• Machine learning applications in healthcare research
• Survival analysis and time-to-event outcomes
• Development of clinical prediction scores

Background:

• Existing time-to-event scores are often ad-hoc, relying on limited variables and clinician knowledge.
• There is a need for robust and efficient methods to generate generic time-to-event scores.
• Interpretability and accessibility are key for clinical utility of scoring systems.

Purpose of the Study:

• To develop and validate AutoScore-Survival, a machine learning-based score generator for time-to-event outcomes with right-censored survival data.
• To provide a robust and user-friendly method for creating interpretable clinical scores.
• To compare the performance of AutoScore-Survival with existing survival models and clinical scores.

Main Methods:

• AutoScore-Survival integrates machine learning for variable selection (Random Survival Forest) and score weighting (Cox regression).
• The method was implemented as an R package for accessibility.
• Performance was evaluated using a 90-day survival prediction study in intensive care unit patients.

Main Results:

• AutoScore-Survival generated more parsimonious scoring systems compared to traditional variable selection methods.
• The performance, measured by integrated area under the curve (0.782), was comparable to other survival models.
• The generated integer-valued time-to-event scores are highly favorable for clinical application due to ease of computation and interpretation.

Conclusions:

• AutoScore-Survival offers a robust, easy-to-use, machine learning-based clinical score generator for time-to-event studies.
• The method provides a systematic approach to facilitate the development of time-to-event scores for clinical practice.
• This tool enhances the interpretability and applicability of survival prediction in healthcare research.