This paper examines the predictive value of peer's observations of an individual, applied to computational models of certain states of that individual. In a study of 28 days, 13 participants provided self-assessments about their level of stress, fatigue, and anxiety, while their smartphone passively recorded the sensor's data. Simultaneously, their designated peers provided assessments about the level of stress, fatigue, and anxiety they perceived from the participant using the PeerMA method. We extracted sensor-derived features (sDFs) from the participant's smartphone, and peer-derived features (pDFs) from the peer's assessments. We evaluated the pDFs on a binary classification task using three machine learning algorithms (Decision Tree-DT, Random Forest-RF, and Extreme Gradient Boosting-XGB). As a result, the classification accuracy consistently increased when the algorithms were trained with the sDFs plus the pDFs, compared the tradition of using only the sDFs. More importantly, the classification accuracy was the highest when we trained the algorithms only with the pDFs (73.3% DT, 73.7% RF, and 71.1% XGB), which represents a unique contribution of this paper. The findings are encouraging about the incorporation of peer's observations in machine learning with potential benefits in the fields of personal sensing and pervasive computing, especially for mental health and well-being.