CPAP telemonitoring is heterogeneous with asynchronous dense nightly device telemetry evidence, episodic patient-report, intervention logs, digital biomarkers with varying sampling rates, missingness mechanisms and clinical meaning. Scalable weekly decisions (who, when, and how) are required by the homecare teams, but that cannot be achieved with a manual review, and the inflexible end-to-end models fail on mixed timestamps and missing modalities. We present a phase-sensitive, multi-model ensemble, which integrates all evidence available to an agreed-upon patient_week representation with an explicit recency and missingness representation and a label (onboarding, optimization, maintenance). Modality specialists (telemetry, PROMs, biomarkers with reliability), task specialists (dropout likelihood, time-to-event horizon and issue flags: leak/residual instability), are trained on their native structure, and their out-of-fold predictions stack affinity-fused to a single calibrated patient-week state. The expert weighting varies with therapy regime and evidence freshness aided by a phase-sensitive gating mechanism to enhance stability. The fused output is translated through an explicit clinician-gated policy mapping (risk tier × horizon × flags × confidence × phase) to predefined action classes and reassessment windows. We assess deployability through time-split validation and rolling monitoring, monitor calibration drift, tier volatility, workload proxies and clinician agreement. This paper puts the concept of personalization into the perspective of a responsive yet manageable orchestration mechanism that allows interpretable and operationally viable decision support across the patient-homecare-physician processes.