| dc.description.abstract | This study evaluates the effects of consuming sungkai leaves (Peronema canescens) on the recovery of COVID-19 patients using Propensity Score Matching (PSM) with machine learning techniques, specifically the Random Forest algorithm, to estimate propensity scores. The research compares three matching methods: Nearest Neighbor Matching, Radius Matching, and Optimal Matching, to identify the best approach for balancing covariates and assessing the impact of sungkai consumption on recovery speed.
The study is based on data collected from COVID-19 survivors in West Sumatra through online questionnaires, comparing those who consumed sungkai leaves with those who did not. PSM is used to balance covariates such as patient age, gender, vaccination status, frequency of infection, isolation setting, and sungkai consumption patterns between treatment and control groups. This approach reduces selection bias and enables robust comparisons.
The analysis reveals that individuals who consumed sungkai leaf remedies exhibit different characteristic distributions compared to non-consumers. Among the matching methods, Optimal Matching is the most effective, achieving the highest Percent Bias Reduction (PBR) of 54.36 percent and providing the best balance of covariates. Re-sampling across 10 iterations confirms the robustness of Optimal Matching, as it consistently produces the highest average PBR values compared to other methods.
Key findings indicate that the duration and method of sungkai consumption significantly influence recovery time. However, other factors such as the number of leaves per brew or daily consumption frequency show no significant relationship with recovery duration. Visualizations of post-matching data reveal substantial improvements in the balance of covariates, confirming the robustness of the matching process.
This study provides evidence that consuming sungkai leaves, particularly with attention to duration and method, may accelerate recovery in COVID-19 patients. Furthermore, it demonstrates the efficacy of combining Random Forest-based propensity score estimation with PSM to evaluate non-randomized clinical data. | |
