This work aims to explore social withdrawal in people with Parkinson's disease (PD), an incurable, neurodegenerative disease that impacts 1\% of people over the age of 60 around the world. PD causes a wide range of motor and non-motor symptoms that significantly influence people's quality of life (QoL). Both motor and non-motor symptoms could cause social withdrawal, and social wellbeing plays a critical role in patients' QoL. Therefore, social withdrawal could be a significant consequence of disease progression and deterioration of QoL. For a disease without a complete cure, the principal aim of the treatment is to improve the QoL of patients. The measurement of Parkinson's progression is the prerequisite of appropriate treatment. However, the clinical assessment of Parkinson's is usually conducted every six months and is based on a snapshot of symptoms which might vary over time. These assessments are done by patients or experts and could be biased by memory or experience. Patients are also aware they are being assessed, which may introduce a Hawthorne effect. Thus, continuous, unobtrusive, and objective measurement is the ambition of Parkinson's monitoring. The smartphone is a popular digital device, and it consumes a significant amount of time for personal, social communications. With embedded sensors, the smartphone can even infer social interactions external to it. Previous studies have confirmed its feasibility for observing people's behaviour without disruption. So, it is a promising tool for unobtrusively tracking social activities, and it fulfils the purpose of a novel monitoring method. Therefore, we initiated a year-long longitudinal study to explore social withdrawal in PD patients. A monitoring application was installed on participants' smartphones to capture all nine potentially social-related data sources, 24 hours a day, seven days a week. Eight standardised clinical/psychological scales for measuring Parkinson's progression, QoL, social withdrawal, and related factors were conducted every two months. Specifically designed diaries were also provided to participants to record their weekly QoL and level of social interaction. With participants joining and dropping out, eight participants finished the whole year of observation. As the continuous monitoring of the smartphone application, more than 10 million raw smartphone data points were obtained from these eight participants. Then twenty-two features were extracted from these raw data to establish personal understandings of the social behaviour of each participant. The COVID-19 pandemic, which significantly impacted people's social lives, occurred during the experiment. But it also provided an opportunity to examine our approach to detecting severe social impact. With the confirmation from the interviews with participants, our method successfully reflected participants' conformance to the government's policies for reducing the transmission of COVID-19 and the intense social deviations caused. For the outcomes of the whole-year study, significant associations were found between clinical/psychological scales and at least one feature for each individual. Our model also achieved at least 0.6 R-squared in numerical prediction and 0.6 F1 scores in direction projection of all participants using multiple linear regression and Naive Bayes methods. Overall, we presented an approach that can adaptively learn the social behaviour of a particular individual and make predictions based on smartphone data. It also shows its strong potential as a reference for clinical/psychological standards. Future work can build upon our efforts to more comprehensive monitoring and a higher validity of the approach. The technique demonstrated in this work could also be applied in wider communities where the patients' social impact needs attention.