Security Risk Assessment Checklist for Dermatology AI Imaging Startups

Kevin Henry

Risk Management

Published Updated

13 minutes read
Share this article
Security Risk Assessment Checklist for Dermatology AI Imaging Startups

A security risk assessment helps you identify where patient images and related data can be exposed, altered, or made unavailable—and decide what to do about each risk. For a dermatology AI imaging startup, the assessment should cover the full image lifecycle, including collection, transfer, storage, annotation, model development, deployment, support, and deletion.

HIPAA does not automatically apply to every health technology startup. Applicability depends on the organization’s role and activities: for example, a startup that creates, receives, maintains, or transmits protected health information on behalf of a covered entity may be a business associate. Start by confirming your status and obligations with qualified counsel; see HHS guidance on covered entities and business associates.

How to map patient image data flows and assessment scope

Define the systems, people, and processes in scope before evaluating threats. If you are subject to the HIPAA Security Rule, its risk analysis covers the confidentiality, integrity, and availability of all electronic protected health information (ePHI) you create, receive, maintain, or transmit. HHS describes this scope in its Guidance on Risk Analysis.

Build a data-flow diagram for each product and use case. Trace images and associated data from the clinic’s camera or device through upload, transfer, cloud storage, annotation, preprocessing, model training or inference, results delivery, customer support, backups, and deletion.

  • Inventory the data: Record clinical images, patient identifiers, dates, encounter details, labels, notes, model inputs and outputs, and any identifiers embedded in filenames or image headers.
  • Inventory the systems: Include endpoints, mobile devices, development and production environments, cloud storage, databases, image-processing services, analytics, logging, backup systems, and vendor platforms.
  • Identify access and transfers: Document who can view, download, annotate, export, or change each data type, and where data crosses organizational or geographic boundaries.
  • Mark retention and deletion points: Include temporary files, cached images, local downloads, logs, backups, and copies used for testing or model development.
  • Record assumptions: Identify whether a dataset is identifiable, de-identified, or subject to a limited data set arrangement, and document the basis for that decision.

Include development and research workflows in scope, not just the live product. A prototype, support ticket, or debugging log can create exposure if it contains a patient image or linkable metadata. HHS says the risk analysis should also account for new technologies and changes to operations when they are planned, rather than waiting until after deployment.

Which threats and vulnerabilities should the assessment cover?

Evaluate threats against confidentiality, integrity, and availability. Consider the likelihood of exploitation and the impact if it occurs, including patient harm, disrupted care, privacy loss, regulatory exposure, and loss of trust. HHS describes risk as involving both the likelihood that a threat exploits a vulnerability and the resulting impact in its risk analysis guidance.

  • Unauthorized access or disclosure: Stolen credentials, overly broad staff access, exposed cloud buckets, misdirected files, compromised vendor accounts, and accidental disclosure through screenshots or support tools.
  • Ransomware and destructive attacks: Malware that encrypts or deletes images, labels, model artifacts, or clinical records; include the effects of delayed access during patient care.
  • Image and metadata manipulation: Image substitution, altered labels, corrupted files, or metadata changes that could mislead clinicians or contaminate training and evaluation data.
  • Application and infrastructure weaknesses: Unpatched software, insecure APIs, weak secrets management, vulnerable dependencies, misconfigured storage, and exposed development environments.
  • Insider and process risks: Inappropriate browsing, bulk exports, accidental sharing, unreviewed production access, and inadequate offboarding.
  • Service disruption: Cloud outages, network failures, unavailable identity services, damaged devices, and loss of a critical vendor or data pipeline.
  • AI-specific data and model risks: Poisoned or mislabeled datasets, unauthorized changes to model weights or thresholds, leakage of training examples, and misuse of model outputs.
  • Physical and environmental events: Lost or stolen devices, power failures, fire, water damage, and other events affecting systems or backups.

For each risk, record the affected asset and data, the relevant threat and vulnerability, existing controls, likelihood, impact, risk rating, mitigation, accountable owner, and due date. A risk list without assigned actions and follow-up does not show how you will manage the risk.

How to protect image privacy and govern AI training data use

Dermatology images should be handled as sensitive health data when they are linked to a patient or can reasonably identify someone in a healthcare context. A full-face photograph or comparable image is specifically listed among identifiers in HHS’s HIPAA de-identification guidance. Skin images may also contain visible identifying features, while filenames, timestamps, clinic details, and associated records can make an otherwise unfamiliar image linkable.

Before using clinical images for research, product improvement, or model training, document the lawful basis and permitted purpose for that use. Do not assume that removing names is enough, that a customer’s agreement automatically authorizes every secondary use, or that pseudonymization makes data de-identified.

  • Choose and document a de-identification method: HHS recognizes Safe Harbor and Expert Determination under the HIPAA Privacy Rule. Confirm that the chosen method fits the data and intended use; see HHS guidance on de-identification of protected health information.
  • Inspect the entire image record: Check pixels, image headers, filenames, captions, annotations, embedded text, and linked tables. Consider faces, tattoos, distinctive marks, and other image details as part of image de-identification and re-identification risk.
  • Limit linkage keys: Store any re-identification key separately, restrict its access, and define when it may be used and by whom.
  • Separate datasets by purpose: Keep clinical operations, annotation, research, and model-development data in appropriately segmented environments with access controls and retention rules.
  • Minimize training data: Use only the fields and images needed for the approved purpose. Remove unnecessary identifiers and restrict exports to approved personnel and destinations.
  • Test for unintended disclosure: Where appropriate, assess whether a model, output, or debugging artifact could reveal training examples or identifying details.
  • Control copies and deletion: Track dataset versions, backups, annotation exports, and derived files; define how corrections, retention limits, or deletion requests are handled.

HHS’s Safe Harbor method requires removal of specified identifiers, including full-face photographs and comparable images, and the Expert Determination method requires an expert to determine that the re-identification risk is very small. Treat de-identification as a documented assessment, not a one-time software operation.

How to secure access, authentication, encryption, and monitoring

Apply role-based least-privilege access so each person and service can reach only the data and functions required for its job. Separate customer, support, annotation, engineering, and production roles; require approval for elevated access; and remove access promptly when a person changes roles or leaves.

  • Use unique user accounts, multi-factor authentication for sensitive access, and managed service identities for automated workflows.
  • Review permissions regularly, especially for production databases, image stores, administrative consoles, exports, and vendor accounts.
  • Protect credentials and API keys in a secrets-management system; restrict their scope, rotate them when exposure is suspected, and avoid placing them in source code or logs.
  • Encrypt image and ePHI transfers over networks and protect stored data with appropriately managed encryption keys.
  • Log access to images and sensitive records, including administrative actions, exports, failed logins, permission changes, and security incidents.
  • Alert on suspicious patterns such as unusual bulk access, repeated failed logins, access from unexpected locations, or attempts to disable logging.
  • Keep audit and security logs protected from alteration, limit their contents to avoid unnecessary patient data, and test that alerts reach someone who can act.

Under the HIPAA Security Rule, encryption is an addressable implementation specification—not a blanket rule that it is either always mandatory or may be ignored. Regulated entities must assess whether it is reasonable and appropriate in their environment; if not, they must document the decision and consider an equivalent alternative measure as applicable. See HHS’s explanation of encryption under the Security Rule and its guidance on addressable implementation specifications.

Ready to assess your HIPAA security risks?

Join thousands of organizations that use Accountable to identify and fix their security gaps.

Take the Free Risk Assessment

How to assess vendors and manage ePHI agreements

List every vendor that can create, receive, maintain, or transmit ePHI, including cloud providers, image-processing services, annotation platforms, analytics tools, customer-support systems, and subcontractors. Confirm each vendor’s role, the data it handles, the locations and systems involved, and what happens when the service ends.

If a vendor is acting as a business associate, the HIPAA Privacy Rule requires a written agreement with specified terms before the covered entity permits it to handle PHI. A business associate that uses a subcontractor to handle PHI must also ensure the applicable restrictions and conditions flow down. Review HHS guidance on business associates and its sample business associate agreement provisions.

Use business associate agreements and related contracts to define permitted data use clearly. In addition to required terms, consider negotiating operational details such as:

  • Which data the vendor may handle and for what purposes, including whether any model training or product-improvement use is permitted.
  • Security safeguards, access restrictions, encryption practices, and subcontractor controls.
  • How quickly and through what channel the vendor must report suspected incidents to your organization.
  • Cooperation with investigations, customer inquiries, and applicable breach-response obligations.
  • Data return or destruction at termination, including copies held by subcontractors, backups, or support systems, subject to applicable retention requirements.
  • How the vendor will communicate material service, security, or subprocessors changes.

Assess evidence as well as contract language. Review security documentation, access and incident processes, independent assessment reports where available, vulnerability handling, backup and recovery capabilities, and the vendor’s ability to meet your requirements. A contract alone does not establish that a vendor’s safeguards are effective.

How to evaluate medical-device cybersecurity and patient safety

Determine whether the product’s intended use and functionality may make it a medical device, and whether the connected product meets the statutory definition of a cyber device. FDA explains that section 524B applies to certain cyber devices included in specified premarket submissions; it is not a universal cybersecurity certification for every health AI startup. Review the FDA cybersecurity in medical devices FAQ and obtain regulatory advice for your product and intended use.

For medical-device threat modeling, connect cybersecurity scenarios to clinical consequences. Assess whether an attacker or failure could change an image, patient association, model version, result, threshold, or clinician-facing message—and whether that change could delay diagnosis or contribute to an incorrect decision. Consider the risks of unavailable results as well as incorrect results.

  • Model threats to image capture, upload, storage, preprocessing, inference, and delivery of results.
  • Protect the integrity and traceability of images, labels, model versions, configuration changes, and clinical outputs.
  • Define how users can recognize uncertainty, missing data, stale results, or service interruptions.
  • Test changes and security updates before release, including their effects on connected systems and clinical workflows.
  • Maintain a vulnerability disclosure process and a plan for vulnerability triage, remediation, communication, and patch validation.

FDA’s current cybersecurity guidance for medical devices includes recommendations for cybersecurity risk management and premarket documentation. For cyber devices subject to section 524B, the statute requires sponsors making covered premarket submissions to provide specified cybersecurity information, including a plan for monitoring, identifying, and addressing vulnerabilities, processes for cybersecurity and updates, and a software bill of materials. See the FDA cybersecurity guidance page and its FAQ on section 524B.

How to document remediation, testing, and continuity plans

Maintain a risk register that links each finding to the affected data or system, likelihood, impact, existing controls, planned remediation, owner, target date, validation evidence, and approval status. Include residual risk documentation: explain what risk remains after controls are applied, why it is accepted or deferred, who approved the decision, and when it will be reviewed.

  • Remediation tracking: Prioritize findings by severity and patient impact, assign accountable owners, set deadlines, and record any compensating controls or accepted exceptions.
  • Security testing: Test applications, APIs, cloud configurations, access rules, dependencies, and relevant device or AI workflows. Confirm that testing is authorized and that findings are retested after fixes.
  • Patch validation: Verify that updates address the identified weakness, do not break essential workflows, and can be deployed to supported environments. Record versions, test results, and exceptions.
  • Incident response: Define who triages reports, isolates affected systems, preserves logs, assesses patient and data impact, communicates with customers, and coordinates any required notification decisions.
  • Continuity and recovery: Maintain protected backups, document restoration priorities, set recovery objectives appropriate to the service, and run restore exercises for images, configuration, models, and critical records.
  • Review and approval: Keep assessment versions, evidence, decisions, test results, and remediation approvals in a controlled location with appropriate access.

HIPAA requires risk analysis and risk management for regulated entities, but HHS does not prescribe one universal assessment interval. Reassess periodically and when meaningful changes—such as new image sources, model uses, vendors, infrastructure, incidents, or newly recognized vulnerabilities—may affect risk. HHS describes risk analysis as ongoing in its risk analysis guidance. In short, map every image-data flow, reduce privacy and security exposure, assess clinical consequences, and document how you will remediate, validate, and monitor each material risk.

FAQs

Are dermatology images considered ePHI when linked to patient metadata?

They may be. When an identifiable image is linked to health information, such as a patient record, diagnosis, or care encounter, it can be protected health information; in electronic form, it may be ePHI. HHS specifically lists full-face photographs and comparable images among identifiers for HIPAA Safe Harbor de-identification. Whether a particular image is identifiable depends on its content, associated data, and the applicable facts. See HHS de-identification guidance.

How should a startup assess re-identification risk in AI training images?

Review the image pixels and all linked or embedded information, including faces, distinctive marks, filenames, timestamps, headers, annotations, and linkage tables. Choose and document a HIPAA de-identification method when applicable, restrict any re-identification key, and assess whether combinations of remaining details could identify a person. HHS describes Safe Harbor and Expert Determination in its de-identification guidance. Removing names alone does not establish that an image set is de-identified.

What should vendor agreements require when vendors handle ePHI?

When a vendor is a business associate, a written business associate agreement must include the applicable HIPAA terms, such as permitted uses and disclosures, safeguards, incident reporting, and subcontractor obligations. Agreements should also clearly address any approved training or secondary use, data return or destruction, and operational incident communications. See HHS guidance on business associates.

How often should a dermatology AI imaging startup reassess cybersecurity risks?

Use a documented, risk-based schedule and reassess when significant changes or new risks arise. HHS describes the HIPAA Security Rule risk analysis as an ongoing process and does not set one universal interval for every organization. Revisit the assessment after material changes to data flows, vendors, models, infrastructure, or operations, and after significant incidents or newly identified vulnerabilities. See HHS’s Guidance on Risk Analysis.

Share this article

Ready to assess your HIPAA security risks?

Join thousands of organizations that use Accountable to identify and fix their security gaps.

Take the Free Risk Assessment

Related Articles