Digital Health, InsurTech and Artificial Intelligence

Introduction

Health insurance is being reshaped by digital distribution, electronic identity, mobile applications, telehealth, connected devices, interoperable health records, cloud computing, automation and artificial intelligence (AI). These technologies do more than replace paper processes. They alter how risk is assessed, how care is accessed, how claims are adjudicated, how customers interact with insurers and how value is created across the health-care ecosystem.

For insurers, digitisation can reduce transaction costs, improve speed and widen access. It can also introduce new risks: exclusion of digitally disadvantaged people, inappropriate automated decisions, cyberattacks, unlawful processing of health data, opaque pricing and discriminatory outcomes. Consequently, digital transformation is not simply an information-technology project. It is a governance, operating-model and public-policy challenge.

This chapter examines the principal technologies used in health insurance and managed care, and develops a governance framework for responsible AI in underwriting, claims, clinical triage and fraud detection.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the role of digital distribution, eKYC and mobile servicing in health insurance.
2. Evaluate telehealth, wearables, electronic health records and interoperability.
3. Describe cloud platforms, robotic process automation and blockchain, including their limitations.
4. Assess AI applications in underwriting, claims, triage and fraud detection.
5. Apply principles of fairness, explainability, privacy, cybersecurity and human oversight.
6. Design an implementation and governance approach for a digitally enabled health insurer.

1. From Digitisation to Digital Health Ecosystems

Digitisation converts analogue information into digital form; for example, scanning a paper claim. Digitalisation uses digital tools to improve an existing process, such as enabling online claim submission. Digital transformation redesigns the process and business model, such as real-time authorisation based on interoperable clinical information.

A health-insurance ecosystem includes policyholders, employers, insurers, intermediaries, hospitals, physicians, pharmacies, laboratories, technology vendors, regulators and public health agencies. Data flows among these actors for eligibility verification, care delivery, payment, utilisation management and quality improvement. The insurer therefore operates within a network rather than as an isolated enterprise.

A useful design principle is minimum necessary data: collect and use only data that are relevant, proportionate and lawful for a stated purpose. More data do not automatically produce better decisions. Poor-quality, unrepresentative or unlawfully obtained data can increase both operational and actuarial risk.

2. Digital Distribution and Customer Service

2.1 Digital distribution

Digital distribution includes insurer websites, comparison platforms, mobile applications, embedded insurance, agents using tablets and application programming interfaces (APIs) connecting insurers to banks, employers or health platforms. It can reduce acquisition costs, provide continuous availability and support personalised communication.

The customer journey commonly includes:

• product discovery and comparison;
• eligibility and needs assessment;
• quotation and premium payment;
• identity verification and consent;
• policy issuance;
• health-service navigation;
• claims submission and settlement; and
• renewal, retention and grievance handling.

The central managerial issue is not merely channel availability but channel suitability. A fully digital channel may disadvantage people with limited connectivity, disabilities, low digital literacy or distrust of online systems. Omnichannel design should allow customers to move between digital, telephone, agent and branch support without losing information or repeating procedures.

2.2 Electronic know-your-customer (eKYC)

eKYC verifies identity through electronic documents, biometric comparison, database checks, digital signatures, one-time passwords or trusted digital identity systems. In health insurance it supports onboarding, beneficiary changes, policy servicing, payment and claims authentication.

A sound eKYC process should include:

1. identity capture and document validation;
2. liveness or possession checks where biometrics are used;
3. sanctions, fraud and duplicate-identity screening where legally authorised;
4. risk-based escalation to human review;
5. consent, notice and retention controls; and
6. an accessible alternative for customers unable to complete digital verification.

Biometrics are sensitive identifiers. Unlike a password, a biometric characteristic cannot easily be replaced after compromise. Storage, encryption, vendor access, retention and deletion must therefore be tightly governed.

2.3 Mobile servicing

Mobile applications can display coverage, provider directories, pre-authorisation status, digital cards, wellness resources, teleconsultation links and claim updates. Push notifications may reduce missed appointments and improve transparency. However, notifications can expose sensitive health information on shared devices. Applications should use privacy-preserving message previews, strong authentication, session controls and clear permission settings.

Important service metrics include first-contact resolution, average response time, abandonment rate, complaint rate, claim-status transparency and accessibility. Speed should not be treated as the only measure of quality: an instantly delivered incorrect denial is worse than a carefully reviewed decision.

3. Telehealth and Virtual Care

Telehealth uses telecommunications to deliver consultation, monitoring, education or clinical coordination at a distance. Modalities include video visits, telephone consultations, asynchronous messaging, remote diagnostics and remote patient monitoring.

For managed care, telehealth may improve access, support chronic-disease management, reduce travel and facilitate early intervention. Insurers may incorporate telehealth into primary-care networks, nurse advice lines, second opinions and post-discharge monitoring.

Its limitations include inadequate physical examination, technology failure, fragmented documentation, unequal broadband access, cross-border licensing issues, inappropriate substitution for in-person care and uncertainty about provider responsibility. Reimbursement policy should distinguish clinically appropriate virtual care from volume-driven overuse.

A telehealth governance framework should specify:

• provider credentialing and jurisdiction;
• informed consent and emergency escalation;
• documentation and record integration;
• prescribing and diagnostic safeguards;
• accessibility and language support;
• data security and recording policy; and
• quality indicators, including avoidable referrals, follow-up completion and patient-reported outcomes.

4. Wearables and Remote Monitoring

Wearables include smart watches, activity trackers, continuous glucose monitors and connected medical devices. They generate data about activity, heart rate, sleep, glucose or other physiological variables. Insurers may use them for wellness programmes, care management, early warning and outcomes measurement.

The distinction between wellness data and clinically validated data is essential. A consumer device may be useful for behavioural feedback without being sufficiently accurate for diagnosis or benefit adjudication. Data may also be missing systematically: older, poorer or sicker individuals may be less likely to own or regularly use a device.

Programmes that link rewards to wearable data require explicit answers to five questions:

1. Is participation genuinely voluntary?
2. What data are collected and for what purpose?
3. Can non-participants obtain equivalent coverage and service?
4. How are inaccurate or missing readings handled?
5. Could the programme indirectly penalise disability, pregnancy, chronic illness or socioeconomic disadvantage?

The insurer should validate devices, document data provenance, monitor representativeness and prohibit secondary use inconsistent with the original notice.

5. Electronic Health Records and Interoperability

An electronic health record (EHR) is a longitudinal digital record of a person’s clinical information. It may contain diagnoses, medications, allergies, laboratory results, imaging reports, procedures and clinical notes. An insurer may access only a legally justified subset, generally for purposes such as care coordination, payment or utilisation review.

Interoperability is the ability of systems to exchange and use information accurately. It has several dimensions:

• technical interoperability: systems can connect and transmit data;
• syntactic interoperability: exchanged data follow a common structure;
• semantic interoperability: terms have consistent meaning; and
• organisational interoperability: institutions align policies, workflows and responsibilities.

Standards such as HL7 FHIR support structured exchange through resources and APIs, while terminologies such as ICD and SNOMED CT support classification and clinical meaning. Standards do not, by themselves, guarantee good interoperability. Mapping errors, incomplete records, incompatible identifiers and inconsistent consent can still produce unsafe results.

Data quality should be evaluated for completeness, validity, consistency, timeliness, uniqueness and provenance. A claim system that receives a diagnosis code without the relevant clinical context may make an apparently precise but substantively incorrect decision.

6. Cloud Platforms, Automation and Blockchain

6.1 Cloud computing

Cloud platforms provide on-demand computing, storage, databases, analytics and security services. They can scale during open-enrolment periods, support disaster recovery and reduce the need for insurers to operate all infrastructure themselves.

Cloud adoption creates shared-responsibility risk. The provider may secure physical infrastructure and core services, while the insurer remains responsible for identity management, configuration, data classification, access privileges, application security and lawful processing. A contract should address location, subcontractors, audit rights, incident notification, portability, service levels, deletion and exit arrangements.

6.2 Automation and robotic process automation

Rule-based automation is effective where inputs are structured and policy rules are stable. Robotic process automation (RPA) can retrieve documents, reconcile payments, populate systems and send routine notifications. It is less suitable for ambiguous clinical narratives or exceptions.

Automation should be accompanied by exception queues, audit logs, segregation of duties, change control and performance monitoring. Automating a poorly designed process merely produces errors faster. A useful principle is straight-through processing with controlled human escalation, rather than automation without accountability.

6.3 Blockchain and distributed ledgers

A distributed ledger records transactions across multiple participants using a consensus mechanism. In health insurance, proposed uses include provider credentialing, consent evidence, claims coordination and audit trails. A ledger can improve tamper evidence and shared visibility where several parties lack a trusted central record.

It is not a universal solution. Health data placed permanently on a ledger may conflict with correction, deletion and minimisation obligations. Storing large clinical files on-chain is generally inefficient and may expose metadata. A safer architecture often stores encrypted data off-chain and places only a hash, pointer or consent event on the ledger. Governance remains essential: who operates nodes, resolves disputes, corrects errors and controls access?

7. Artificial Intelligence in Health Insurance

AI refers broadly to computational systems performing tasks associated with human intelligence. Machine learning (ML) systems infer patterns from data; deep learning uses multilayer neural networks; natural-language processing handles text and speech; generative AI produces new text, images or code.

The predictive value of a model depends on the population, outcome definition, data-generating process and decision context. A high-performing model in one hospital or product line may fail after deployment because of population shift, coding changes or altered clinical practice.

7.1 Underwriting and pricing

AI can assist risk classification, document extraction, eligibility checks and product recommendation. In health insurance, legal restrictions may limit the use of health information, genetic information, protected characteristics or proxies for them. Even when a variable is technically predictive, its use may be socially unacceptable or unlawful.

Traditional actuarial governance remains necessary. Models should be compared with established rating methods, calibrated on relevant populations and tested for adverse impact. Human review is especially important where a decision affects access, price or coverage.

7.2 Claims management

AI can classify claims, extract information from invoices, detect missing documents, estimate complexity and prioritise cases. It can assist, but should not silently replace, substantive claims adjudication. Denials require a documented policy basis, evidence trail, explanation and appeal route.

Generative AI introduces additional risks, including hallucinated policy interpretations, leakage of confidential data, unsupported citations and inconsistent outputs. It should initially be limited to bounded tasks such as summarisation with source links, subject to review.

7.3 Clinical triage and care navigation

Triage tools estimate urgency or route members to appropriate services. They should be designed as decision support, not as a substitute for emergency services or professional judgment. Safety thresholds should favour escalation when uncertainty or potential severity is high. Testing must include rare but serious conditions, language variation, disability and atypical presentations.

7.4 Fraud, waste and abuse detection

Anomaly detection can identify unusual billing patterns, duplicate services, improbable provider behaviour or networks of related claims. A flagged claim is not proof of fraud. Investigations must distinguish error, legitimate clinical complexity and intentional deception. Excessive false positives can delay legitimate payment and unfairly stigmatise providers or members.

8. Model Evaluation, Fairness and Explainability

A classification model produces outcomes such as “investigate” or “do not investigate.” Common measures include:

Here, TP is true positive, FP false positive, FN false negative and TN true negative. Precision answers: among flagged cases, how many truly belong to the target class? Recall answers: among all target cases, how many were identified? The measures are meaningful only when the target definition and decision threshold are clear.

Accuracy can be misleading with imbalanced data. For example, a system that labels every claim “non-fraudulent” may appear accurate when fraud is rare but has zero fraud recall.

Fairness is multidimensional. Relevant assessments include:

• demographic parity: comparable positive decision rates across groups;
• equal opportunity: comparable true-positive rates across groups;
• equalised odds: comparable true-positive and false-positive rates; and
• calibration: comparable meaning of a predicted risk score across groups.

These criteria can conflict, particularly where base rates differ. Therefore, fairness is not solved by selecting one universal metric. The organisation must define the legally and ethically relevant harm, involve affected stakeholders and document trade-offs.

Explainability operates at several levels: global model behaviour, case-specific reasons, data provenance, decision rules and counterfactual information. A customer generally needs a usable explanation of the principal factors and how to challenge an outcome, not merely a technical description of model architecture.

9. Privacy, Cybersecurity and Responsible Governance

Health data are highly sensitive. Privacy governance should cover purpose limitation, lawful basis, informed notice, consent where required, access and correction rights, retention limits, data minimisation, de-identification and controlled secondary use. De-identification reduces risk but does not guarantee anonymity, especially when datasets can be linked.

Cybersecurity controls should include:

• least-privilege access and multifactor authentication;
• encryption in transit and at rest;
• network segmentation and endpoint protection;
• secure software development and vulnerability management;
• immutable or protected backups;
• logging, monitoring and anomaly detection;
• vendor due diligence and incident obligations; and
• tested incident-response and business-continuity plans.

AI governance should begin before procurement. A practical model lifecycle is:

1. define the purpose, affected people and prohibited uses;
2. assess legal, clinical, actuarial and ethical risks;
3. document data sources, consent, quality and representativeness;
4. develop and validate the model using appropriate benchmarks;
5. conduct fairness, robustness, privacy and security testing;
6. obtain independent approval before deployment;
7. monitor drift, complaints, overrides, subgroup outcomes and incidents; and
8. retire or retrain the model when performance or legitimacy deteriorates.

Accountability must be assigned to named owners. A vendor cannot assume the insurer’s regulatory responsibility. Boards and senior management should receive understandable reports on material automated decisions, incidents, model limitations and remediation.

10. Implementation Architecture and Performance Measurement

A mature digital health-insurance architecture separates channels, integration, data, decision services and governance. APIs connect applications to policy, provider, claims and payment systems. An enterprise identity layer manages authentication and authorisation. A data platform supports approved analytics, while a model-management layer controls versioning, validation and monitoring.

Implementation should proceed incrementally: establish data governance, improve interfaces, pilot a low-risk use case, evaluate outcomes, then scale. High-impact decisions require stronger controls than administrative automation.

A balanced performance dashboard may include:

A business case should include total cost of ownership, integration, training, monitoring, compliance and exit costs—not just software licence fees. Benefits should be measured against a credible baseline and assessed for unintended effects.

11. Illustrative Case: CareConnect Health Insurer

Background

CareConnect, a fictional regional insurer, receives a high volume of outpatient claims and has long telephone queues. It proposes a mobile application, telehealth partnership and AI-assisted claims triage.

Problem

The insurer wants faster settlement and better navigation, but its historical claims data contain inconsistent provider codes. Older members have lower application use, and previous automated fraud rules produced disproportionately high investigation rates for clinics serving rural communities.

Analysis

A technology-first rollout would reproduce data-quality and access problems. The insurer therefore conducts a data inventory, reviews legal authority for each data element, tests provider-code mappings and analyses error rates by geography, age and language. It also distinguishes administrative prioritisation from an adverse claim decision.

Approach and solution

CareConnect deploys digital cards and claim-status tracking while retaining telephone and agent channels. Telehealth is integrated with the member record only after consent and provider verification. The AI model prioritises claims for human review rather than denying them. Its evaluation reports precision, recall and subgroup false-positive rates. A rural-provider advisory group reviews flagged patterns.

The model card documents purpose, training period, exclusions, intended users, known limitations, thresholds and escalation rules. Every adverse decision remains subject to human review, reasons, correction and appeal. Security testing covers the mobile application, APIs and vendors.

Outcome

In the first evaluation period, average administrative processing time falls, but the insurer discovers that the model performs poorly for a newly contracted provider network. It pauses expansion, corrects coding and retrains the model. The case demonstrates that responsible deployment is iterative: operational improvement is valuable, but monitoring can reveal that apparent efficiency conceals unequal or unreliable performance.

Learning points

1. Digital access must remain inclusive and multi-channel.
2. Data quality is a prerequisite for trustworthy AI.
3. A fraud flag is an investigative signal, not a finding.
4. Fairness monitoring must continue after deployment.
5. Human accountability and appeal rights are operational controls, not merely ethical statements.

Contemporary Developments and Strategic Implications

Generative AI copilots, synthetic data, digital therapeutics, hospital-at-home programmes, privacy-enhancing computation and real-time health APIs are expanding the digital-health frontier. Their value depends on evidence, integration and governance. Synthetic data can support development but may fail to preserve rare cases or eliminate re-identification risk. Federated or privacy-enhancing approaches can reduce data movement but add technical complexity.

Executives should ask whether a proposed innovation improves health outcomes, affordability, access or trust; whether the organisation can explain and contest its decisions; and whether it can safely stop the system. Strategic advantage increasingly comes from reliable data foundations, interoperable partnerships and institutional trust rather than from deploying the largest model.

Chapter Summary

Digital health and InsurTech transform distribution, identity verification, servicing, care delivery, records, payment and risk management. Telehealth and wearables can extend access and support prevention, but require clinical validation and inclusion safeguards. EHR interoperability depends on technical, syntactic, semantic and organisational alignment. Cloud, automation and blockchain offer operational advantages but create shared-responsibility, governance and privacy challenges.

AI can support underwriting, claims, triage and fraud detection, yet predictive performance alone is insufficient. Precision, recall, calibration, fairness, explainability, privacy, cybersecurity, human review and appeal rights must be considered together. Responsible deployment requires lifecycle governance, data-quality controls, subgroup monitoring, accountable ownership and the ability to suspend or retire unsafe systems.

Key Terms

Digital transformation: Redesign of processes and business models using digital capabilities.

eKYC: Electronic verification of a customer’s identity.

Telehealth: Delivery of health services through telecommunications.

Wearable: A connected device worn or used to collect behavioural or physiological data.

Electronic health record: A longitudinal digital record of clinical information.

Interoperability: The ability of systems to exchange and use information accurately.

API: A defined interface that enables software systems to communicate.

Cloud computing: On-demand access to shared computing resources.

RPA: Software automation of structured, repetitive tasks.

Blockchain: A distributed ledger using agreed mechanisms to record transactions.

Machine learning: Methods that infer patterns from data for prediction or decision support.

Model drift: Change in data or relationships that reduces model performance over time.

Precision: The proportion of positive predictions that are true positives.

Recall: The proportion of actual positive cases that are identified.

Calibration: Agreement between predicted probabilities and observed frequencies.

Explainability: The extent to which a system’s operation or outcome can be understood.

Data minimisation: Limiting collection and use to what is necessary for a stated purpose.

Human-in-the-loop: Meaningful human oversight within an automated decision process.

Federated learning: Model training across distributed data without centrally pooling the raw data.

Model card: Documentation describing a model’s purpose, performance, limitations and use conditions.

Review Questions

Short Answer Questions

1. Distinguish digitisation, digitalisation and digital transformation.
2. What are the four dimensions of interoperability?
3. Why should a fraud flag not be treated as proof of fraud?
4. Define precision and recall.
5. State two cybersecurity controls relevant to a mobile health-insurance application.

Long Answer Questions

1. Evaluate the benefits and risks of telehealth for a managed-care organisation.
2. Explain how an insurer should govern wearable-device data in a wellness programme.
3. Discuss the advantages and limitations of cloud computing, RPA and blockchain in health insurance.
4. Explain why fairness criteria may conflict in AI-based insurance decisions.
5. Develop a lifecycle governance framework for an AI claims system.

Analytical / Application Questions

1. An insurer’s fraud model flags 1,000 claims, of which 250 are confirmed fraudulent. It misses 100 fraudulent claims that were not flagged. Calculate precision and recall, and explain what each measure means for investigation capacity and leakage.
2. A telehealth programme has high utilisation but rising emergency referrals and low uptake among older members. Diagnose possible causes and propose a balanced evaluation plan.
3. Design a governance response after discovering that an automated underwriting tool produces materially higher adverse decisions for a demographic group, even though overall accuracy has improved.

Multiple Choice Questions with Answers

Question 1

Which dimension of interoperability concerns whether exchanged terms have consistent meaning?

A. Technical interoperability

B. Semantic interoperability

C. Organisational interoperability

D. Financial interoperability

Correct Answer: B. Semantic interoperability

Question 2

In a health-insurance AI system, a fraud flag should generally be treated as:

A. Automatic proof of fraud

B. A reason to cancel the policy immediately

C. A signal for proportionate investigation

D. A replacement for claims documentation

Correct Answer: C. A signal for proportionate investigation

Question 3

Which measure answers the question: among all actual positive cases, how many were identified?

A. Precision

B. Recall

C. Specificity

D. Calibration

Correct Answer: B. Recall

Question 4

What is a principal limitation of storing clinical information directly on an immutable ledger?

A. It prevents any audit trail

B. It may conflict with correction and deletion obligations

C. It eliminates the need for access controls

D. It guarantees semantic interoperability

Correct Answer: B. It may conflict with correction and deletion obligations

Question 5

Which is an example of syntactic interoperability?

A. Agreement on the meaning of a diagnosis

B. Agreement on organisational accountability

C. Use of a common data structure for exchanged records

D. Agreement on reimbursement policy

Correct Answer: C. Use of a common data structure for exchanged records

Question 6

Why can accuracy be misleading when evaluating fraud detection?

A. Accuracy cannot be calculated for binary outcomes

B. Fraud datasets may be imbalanced, allowing a trivial majority prediction to appear strong

C. Accuracy measures only privacy risk

D. Accuracy is identical to recall in every dataset

Correct Answer: B. Fraud datasets may be imbalanced, allowing a trivial majority prediction to appear strong

Question 7

Which practice best reflects data minimisation?

A. Collecting every available wearable variable indefinitely

B. Collecting only data relevant and necessary for a stated purpose

C. Sharing raw health data with all vendors

D. Retaining data because storage is inexpensive

Correct Answer: B. Collecting only data relevant and necessary for a stated purpose

Question 8

In the illustrative CareConnect case, the AI model was designed primarily to:

A. Automatically deny suspicious claims

B. Prioritise claims for human review

C. Replace provider credentialing

D. Set premiums without actuarial review

Correct Answer: B. Prioritise claims for human review

Question 9

Which control is most directly associated with least privilege?

A. Giving every employee administrator access

B. Allowing users only the access required for their role

C. Disabling audit logs

D. Storing passwords in application code

Correct Answer: B. Allowing users only the access required for their role

Question 10

A model flags 1,000 claims and 250 are confirmed fraudulent. What is its precision?

A. 10%

B. 25%

C. 40%

D. 75%

Correct Answer: B. 25%