Adaptive deep brain stimulation (aDBS) offers unprecedented precision in the treatment of Parkinson's disease. Current aDBS control algorithms depend on brain signal biomarkers, such as basal ganglia oscillatory activity in the beta (8-30 Hz) range. Even after extensive optimization through a specialized medical team, about one third of patients may remain ineligible for aDBS due to insufficient biomarker fidelity. Moreover, while the marker broadly correlates with disease severity, it does not recognize how specific symptoms independently fluctuate over time. A concept to address these shortcomings is to rely on machine learning based brain signal decoding. Here, we trained decoders on over 500 hours of invasive multisite recordings from cortex and deep brain targets to predict the wearable based estimates of PD symptoms and side-effects without patient individual training. Recordings were streamed from brain implants while patients were at home on their usual antiparkinsonian treatment. The resulting models robustly outperformed individually defined beta activity, while providing three major conceptual advances: they performed even in patients without a basal ganglia beta rhythm, generalized across patients without requiring retraining or adaptation and robustly and differentially and simultaneously decoded the severity of bradykinesia, tremor and dyskinesia. In a proof-of-concept simulation, we demonstrate how these models could be used to steer stimulation fields dynamically to predefined symptom-response brain networks. This fusion between adaptive and connectomic DBS may define when to stimulate which brain circuit for optimal symptom-specific control in real-time and pave the way for a new generation of fully automatized symptom-specific neuromodulation approaches.