Definition, custody objects, and neighboring service boundaries
Sleep data asset compliance custody services place sleep data and the controls surrounding it in a verifiable environment under the data provider's documented instructions, an analysis of applicable requirements, and a contract. They continuously govern access, use, retention, delivery, and exit and sit within WSIC 6.4.3, Data Asset Operations. Custodied assets may include PSG, HSAT, EEG, EOG, EMG, ECG, oximetry, PPG, activity, radar, audio, video, questionnaires, clinical covariates, and device logs. They also include metadata, labels, dictionaries, provenance, versions, consent or authorization evidence, use restrictions, access policies, cryptographic keys, audit records, and derived copies. The service is valuable when it keeps the answer to who may do what to which version under which conditions enforceable and reviewable, not merely when it stores files in a cloud bucket.
Custody does not automatically transfer ownership, intellectual property, controller accountability, or clinical responsibility. Nor is it synonymous with establishing title, data quality rating, de-identification, backup, disaster recovery, a data sharing platform, a trusted research environment, blockchain notarization, or data brokerage. Some of those capabilities may be components, but each needs its own scope and acceptance criteria. If a provider selects a new purpose, reuses data without authority, or brokers access on its own initiative, its role and obligations may change; it cannot rely indefinitely on the label of passive custodian. Compliance in this entry means operating controls for identified jurisdictions, parties, purposes, and periods. It is not a permanent warranty against every legal requirement or every possible scenario.
Roles, intake inventory, and lawful-use boundaries
At initiation, the parties should map the data provider, controller or accountable organization, processor, custody provider, cloud and subprocessors, approved researchers, clinical organizations, security and privacy officers, key custodians, and incident contacts. The matrix should identify who approves purposes, verifies requesters, can decrypt, reviews outputs, responds to individual rights requests, and coordinates required notices. The provider processes only confirmed instructions. Conflicting instructions, unclear authority, out-of-scope demands, or apparently unlawful activity should pause the affected operation and enter the agreed escalation process. A customer instruction is not by itself a defense for executing an operation that the provider knows is unauthorized.
Before intake, the custodian needs an asset manifest and admission record at the dataset, batch, or object-version level. It should cover provenance, collecting institution, modality, format, volume, sensitivity, identifiability, authority, consent or waiver, allowed and prohibited purposes, population, territorial and transfer restrictions, retention, legal holds, withdrawal handling, and downstream sharing. File manifests, checksums, dictionaries, de-identification status, and transport integrity should be reconciled, while unknown, damaged, or unauthorized material is quarantined. Technical ability to ingest data is not authority to possess it. Removing direct identifiers does not automatically eliminate linkage risk, group harm, or contractual restrictions.
Segmented architecture, key management, and privileged access
The custody architecture should be segmented by sensitivity, purpose, tenant, project, and lifecycle. Intake quarantine, preserved originals, analytic workspaces, approved release areas, backup, and audit zones require controlled network paths, copy directions, and management planes. Encryption in transit and at rest is only a baseline. The design must address key generation, possession, rotation, revocation, backup, recovery, and staff departure, with suitable separation between keys and ciphertext, operations and approval, and development and production. Customer-held keys, hardware protection, or dual control may be appropriate, but the threat model, recovery need, and contract determine that choice; a product name does not establish the resulting assurance level.
Human and service accounts need unique identity, least privilege, separation of duties, strong authentication, time-limited just-in-time access, and recurring review. Privileged operations, emergency break-glass access, key use, policy changes, bulk download, and cross-zone copying require additional approval, alerting, and retrospective review. Shared administrator accounts obscure accountability and should not be the normal path. Devices, APIs, tokens, scheduled jobs, and machine identities belong in the same inventory and revocation process. A sleep laboratory, remote collection platform, algorithm team, and external researcher do not receive implicit access merely because they share a corporate parent. Departure, project closure, purpose expiry, or risk change must promptly remove access.
Authorized use, controlled outputs, and traceable audit
Every access request should bind an identified requester and institution to a purpose, data scope and version, method, environment, duration, requested outputs, and redisclosure limits. An authorized decision maker then grants only the minimum required access. High-risk data may call for a controlled workspace, remote query, de-identified view, field masking, query limits, or synthetic test data. Unapproved copies must not move to personal devices, public collaboration areas, third-party generative AI, or unreviewed software services. Output review should consider small cells, free text, recognizable faces, audio and video, rare condition combinations, precise time and location, linkable identifiers, and model memorization while preserving information genuinely needed for reproducibility.
The audit chain should show who, when, under which approval, performed a view, query, export, deletion, decryption, policy change, or configuration change on a specific version and where the result went. Logs need consistent time sources, proportionate tamper resistance, access isolation, retention, and anomaly detection, with investigation routes for bulk transfer, failed authentication, off-hours access, privilege escalation, and unusual queries. Logs can themselves contain personal data and security secrets and should not be universally visible. Checksums, version manifests, and signed records support integrity, but not every custody arrangement requires a blockchain. A technical timestamp alone does not prove that authorization was lawful, content was true, or the use was legitimate.
Security operations, incident response, and continuity
Ongoing operations should maintain an asset inventory and data-flow map grounded in risk assessment. The program should cover secure configuration, vulnerabilities and patches, malware, loss prevention, endpoint and network monitoring, personnel screening and training, supply-chain changes, penetration or control testing, and tracked remediation. Controls must support governance, identification, protection, detection, response, and recovery instead of treating one successful audit as an endpoint. For ransomware, credential compromise, insider misuse, accidental publication, cross-tenant access, or unauthorized export, the provider preserves evidence, contains impact, determines affected data and people, coordinates notices under the contract and applicable rules, and records causes, remedies, and recovery decisions.
Primary storage, version history, backup, and disaster-recovery copies serve different objectives. Each needs documented coverage, encryption, isolation, retention, recovery authority, and propagation of deletion. Backup should not share every production credential and failure domain, and scheduled restoration exercises must demonstrate usability; a successful backup job does not prove complete restoration. Recovery point and recovery time objectives, availability commitments, maximum tolerable disruption, and manual fallback should be defined and tested according to business risk. No provider can honestly promise that service will never fail or data will never leak. Mature custody makes failures discoverable, contains their effect, and produces verifiable recovery.
Contracts, subprocessors, evidence, and acceptance
Contracts and data-processing schedules should match the deployed architecture. They define roles, instructions, purpose and scope, locations and transfer routes, subprocessors, confidentiality, workforce access, technical and organizational measures, audit evidence, incident timing, support for rights requests, continuity, change control, return and deletion, allocation of responsibility, and dispute handling. A subprocessor must not become an invisible break in accountability. A new region, cloud service, support workforce, or analytic tool should trigger notice, assessment, or renewed approval according to risk. Where a health-information regime requires a business-associate or comparable arrangement, the parties determine applicability and flow obligations down instead of mechanically applying one jurisdiction's template to all sleep data.
Acceptance can inspect asset and access inventories, configuration baselines, key and access reviews, log samples, incident exercises, restoration tests, vulnerability remediation, subprocessor lists, deletion evidence, and independent assessment reports. Claims must be matched to the exact legal entity, system, region, period, and control scope assessed. A certificate, audit opinion, or cloud compliance badge is scoped evidence, not a substitute for the customer's risk analysis and not proof of flawless future operation. Pricing should disclose storage, query, log, recovery exercise, network egress, emergency support, migration, and deletion charges so that proprietary formats, inaccessible keys, or opaque exit costs do not create involuntary lock-in.
Change, return and deletion, and claim boundaries
Changes in data, purpose, personnel, jurisdiction, systems, keys, subprocessors, or risk require updated inventories and flows and a decision on renewed authorization, additional notice, modified controls, or suspension. The exit plan should be settled when contracting. The provider returns data, metadata, version manifests, necessary logs, and portable configuration in agreed formats, verifies completeness, then revokes accounts, tokens, interfaces, and key access. It also follows an executable deletion schedule for primary and secondary copies, caches, temporary files, search indexes, test copies, and backups. A legal hold, backup that cannot be immediately erased, or other exception needs a documented basis, isolated access, and resumed deletion when the exception ends, rather than a vague promise of instant and total erasure.
Successful custody acceptance does not prove source accuracy, label reliability, perpetual authorization, or fair and effective algorithms. It is not a clinical diagnosis, medical-device clearance, legal opinion, or confirmation of asset value. Marketing should not turn encryption, de-identification, blockchain, zero trust, a certification, or on-premises deployment into claims of absolute compliance, zero risk, or impossibility of breach. Custodied data must not be used for advertising, profiling, model training, or product development without explicit authority. A responsible service describes scope, auditable evidence, residual risk, exceptions, and exit capability and leaves continuing governance and final judgment with the accountable data holder.