Current blood pressure (BP) estimation methods have not achieved an accurate and adaptable approach for ambulatory diagnosis and monitoring applications of populations at risk of cardiovascular disease, generally due to a limited sample size. This paper introduces an algorithm for BP estimation solely reliant on photoplethysmography (PPG) signals and demographic features. It automatically obtains signal features and employs the Markov Blanket (MB) feature selection to discern informative and transmissible features, achieving a robust space adaptable to the population shift. This approach was validated with the Aurora-BP database, compromising ambulatory wearable cuffless BP measurements for over 500 individuals. After evaluating several machine-learning regression methods, Gradient Boosting emerged as the most effective. According to the MB feature selection, temporal, frequency, and demographic features ranked highest in importance, while statistical ones were deemed non-significant. A comparative assessment of a generic model (trained on unclassified BP data) and specialized models (tailored to each distinct BP population), demonstrated a consistent superiority of our proposed MB feature space with a mean absolute error of 10.2 mmHg (0.28) for systolic BP and 6.7 mmHg (0.18) for diastolic BP on the whole dataset. Moreover, we present a first comparison of in-clinic vs. ambulatory models, with performance significantly lower for the latter with a drop of 2.85 mmHg in systolic ( ) and 2.82 mmHg for diastolic ( ) estimation errors. This work contributes to the resilient understanding of BP estimation algorithms from PPG signals, providing causal features in the signal and quantifying the disparities between ambulatory and in-clinic measurements.