Ambient Assisted Living (AAL) systems are becoming increasingly important for providing personalised assistance in smart homes. One key component for such systems is detecting and localising humans in different areas of the home, which can enhance contextual information to provide efficient support to the human user. Recent approaches often lack interpretability and compromise user privacy. This work introduces an interpretable, room-level human presence detection system that relies solely on low-cost, privacy-conserving ambient sensors typically used in smart homes. We have developed and evaluated a solution for presence detection based on data collected from a single participant in the Robot House, an ambient assisted living space at the University of Hertfordshire. We developed two models to perform this task, a Random Forest (RF) model and a more complex Long Short-Term Memory (LSTM) model across a triad of test scenarios, including full sensor set, sensor dropout and room dropout. We tested the performance of both models using conventional train-test splits and on an entirely unseen data to assess the generalisation. While LSTM achieved comparable results, RF performed better on new, unseen data, with an accuracy of 91.43% vs. 62.69% for RF and LSTM, respectively. The RF also achieved comparative results against two state-of-the-art models, HOOD and CSI-BiLSTM, with the advantages of being easy to interpret and working better in situations where privacy and cost are important. Overall, our work provides the basis for creating a scalable and interpretable solution for finding a person’s location in smart homes.