Digital Quality, Health Information and Artificial Intelligence

Introduction and context

Healthcare quality has become inseparable from the quality of information used to design, deliver, monitor and improve care. Electronic health records (EHRs), clinical decision support (CDS), telehealth, interoperability platforms, dashboards, robotic process automation and artificial intelligence (AI) can improve access, safety, continuity and efficiency. They can also create new hazards: incorrect data may be propagated rapidly, alerts may be ignored, poorly designed interfaces may encourage unsafe workarounds, and biased algorithms may worsen inequity.

Digital quality therefore means more than purchasing information technology. It is the disciplined management of digital systems so that information is accurate, available, timely, secure, understandable and used in ways that improve patient outcomes. A hospital manager must evaluate both technical performance and effects on clinical work, patient experience, staff workload, equity and safety.

This chapter treats digital systems as sociotechnical systems. Their performance depends on the interaction of software, hardware, data, people, clinical processes, organisational culture, governance and external regulation.

Learning objectives

By the end of this chapter, the learner should be able to:

1. Explain the quality and safety functions of EHRs, CDS, telehealth and interoperability.
2. Assess data quality using dimensions such as accuracy, completeness, timeliness and validity.
3. Interpret digital dashboards and distinguish useful indicators from misleading measures.
4. Identify cybersecurity, privacy and technology-related clinical risks.
5. Explain automation, machine learning, algorithmic bias and human oversight.
6. Apply a structured approach to AI validation, implementation and monitoring.
7. Recommend governance controls for safe and equitable digital healthcare.

1. Digital health as a quality system

A digital health system performs several connected functions:

• Capture: recording observations, diagnoses, orders, results and patient-reported information.
• Exchange: transferring information between departments, organisations and patients.
• Interpretation: transforming data into alerts, risk scores, trends or recommendations.
• Action: supporting clinical, operational and managerial decisions.
• Learning: using outcomes and experience to improve processes.

A useful quality chain is:

Data → Information → Knowledge → Decision → Action → Outcome → Learning

Failure at any point can undermine safety. For example, a medication allergy entered incorrectly is poor data; displaying it prominently is accurate transmission of inaccurate information; prescribing on the basis of it is an unsafe decision. Conversely, correct data that arrive after a clinical decision may be functionally useless.

Digital quality should be assessed through six questions:

1. Is the information correct?
2. Is it available to the right person at the right time?
3. Is it understandable and relevant to the task?
4. Is the system usable under real working conditions?
5. Is access lawful, secure and appropriately limited?
6. Does use of the system improve outcomes without creating disproportionate burden or inequity?

2. Electronic health records

An EHR is a longitudinal digital record containing health information generated across episodes and settings of care. Typical components include demographics, encounters, diagnoses, medications, allergies, laboratory results, imaging reports, nursing documentation, care plans, discharge summaries and billing or administrative data.

Quality benefits

EHRs can improve quality by making information legible and available, supporting medication reconciliation, reducing duplicate tests, enabling structured documentation and providing data for audit. They can also facilitate care coordination through shared care plans and automatic notification of abnormal results.

Safety risks

EHR-related hazards include:

• wrong-patient selection;
• copy-and-paste propagation of outdated information;
• incomplete medication reconciliation;
• incorrect default values;
• confusing screens and hidden information;
• alert overload;
• delayed interface messages;
• downtime and loss of access;
• excessive documentation burden;
• inappropriate access to confidential records.

A technically correct EHR can still be unsafe if its workflow conflicts with clinical practice. Usability testing should therefore involve representative users performing realistic tasks, including interruptions, high workload and emergency conditions.

Evaluation framework

Managers should evaluate an EHR across four levels:

Availability is commonly expressed as:

Availability is meaningful only when the measurement period and definition of downtime are specified. A system can have high availability but still cause harm through incorrect data or poor usability.

3. Clinical decision support

Clinical decision support provides patient-specific information, assessments or recommendations to assist healthcare decisions. Examples include drug–drug interaction warnings, renal-dose prompts, preventive-care reminders, sepsis screening, order sets and evidence links.

CDS may be:

• Passive, such as a reference page;
• Informational, such as a laboratory trend;
• Advisory, such as a dose recommendation;
• Interruptive, such as a pop-up alert;
• Protocol-driven, such as an order set or pathway.

The central principle is that CDS supports, rather than replaces, professional judgement. A recommendation must be interpretable, relevant to the patient and presented at a point where action is possible.

Alert fatigue and the five rights

Alert fatigue occurs when frequent or low-value alerts reduce attention to important warnings. Effective CDS follows the “five rights”:

1. the right information;
2. to the right person;
3. in the right format;
4. through the right channel;
5. at the right time in workflow.

Evaluation should include override rates, clinically appropriate overrides, response time, adverse events and user experience. A high override rate is not automatically evidence of failure; it may indicate poor specificity, inappropriate thresholds or legitimate clinical exceptions.

4. Interoperability and health information exchange

Interoperability is the ability of systems to exchange and use information consistently. It has three related dimensions:

• Technical interoperability: systems can transmit data.
• Syntactic interoperability: exchanged data follow a common structure or format.
• Semantic interoperability: receiving systems assign the same meaning to the data.

A laboratory result transmitted as “positive” but without the test name, specimen type, units or reference range may be technically exchanged but clinically unsafe. Interoperability requires common identifiers, terminology, metadata, patient matching and governance.

Health information exchange can improve referrals, reduce repeated investigations and support transitions of care. Risks include mismatched patient identities, uncertain provenance, conflicting versions of information and inappropriate secondary use.

Managers should require:

• positive patient identification and duplicate-record controls;
• documented data provenance;
• standard terminology and units;
• reconciliation of conflicting data;
• role-based access;
• interface monitoring and exception queues;
• downtime and recovery procedures.

5. Telehealth and virtual care

Telehealth includes video consultations, telephone care, remote monitoring, asynchronous messaging and store-and-forward services. It may improve access for rural populations, people with mobility limitations and patients needing follow-up rather than physical examination.

A safe telehealth pathway begins with clinical appropriateness and includes identity verification, consent, privacy, reliable connectivity, emergency location, escalation arrangements and documentation. Clinicians must recognise when remote assessment is insufficient. Chest pain, severe breathlessness, altered consciousness or safeguarding concerns may require immediate in-person evaluation.

Quality indicators may include waiting time, completion rate, conversion to in-person care, avoidable emergency attendance, patient experience, technical failure and outcomes stratified by age, language, disability, socioeconomic status and connectivity. Digital access is not automatically equitable: a virtual-only model may exclude people without devices, data, private space or digital literacy.

6. Digital dashboards and measurement

A dashboard converts selected data into visual information for monitoring and action. It should answer a management question, identify variation and make responsibility clear. Attractive graphics are not evidence of quality.

A sound dashboard specifies:

• the indicator definition and numerator/denominator;
• data source and extraction date;
• target or benchmark;
• reporting frequency;
• responsible owner;
• stratification variables;
• rules for missing and revised data;
• action thresholds.

For a rate:

The multiplier may be 100, 1,000 or 100,000 depending on the event and convention. The numerator must be a subset of the defined denominator. For example, a falls rate per 1,000 occupied bed-days is:

This measure is not directly comparable with a rate per 1,000 admissions. Managers must distinguish signal from noise, account for case mix and avoid interpreting correlation as causation. Dashboards should display uncertainty or at least provide context when volumes are small.

A balanced digital quality dashboard may combine outcome, process, balancing and equity measures. For example, faster discharge is not necessarily improvement if readmissions, medication errors or patient complaints increase.

7. Automation and robotic process automation

Automation uses software or machines to perform tasks with limited human intervention. In hospitals it may support appointment reminders, claims processing, stock management, coding, document classification, result routing and repetitive data entry.

Automation is most suitable where rules are stable, inputs are reliable, exceptions are identifiable and consequences are reversible. It is less suitable where data are ambiguous, decisions are ethically sensitive or the cost of an undetected error is high.

A safe automation design includes:

1. a clearly defined purpose;
2. documented input and output variables;
3. validation against real cases;
4. exception handling;
5. audit logs;
6. a human escalation route;
7. rollback or shutdown capability;
8. post-deployment monitoring.

Automation bias occurs when users accept a computer-generated recommendation without adequate independent assessment. Human involvement is meaningful only when the person has time, competence, authority and sufficient information to challenge the system.

8. Artificial intelligence in healthcare

AI refers broadly to computational systems that perform tasks associated with human cognitive functions, such as classification, prediction, language processing or pattern recognition. Machine learning systems learn relationships from data rather than relying solely on explicitly programmed rules.

A predictive model may estimate a probability:

Here, Y = 1 represents occurrence of the outcome and X represents patient or process features. A probability is not a diagnosis or a command. Its clinical meaning depends on calibration, threshold selection, prevalence, consequences of false positives and consequences of false negatives.

Common performance concepts

• Sensitivity: proportion of people with the condition who are identified.
• Specificity: proportion of people without the condition who are correctly classified as negative.
• Positive predictive value: proportion of positive classifications that are true positives.
• Negative predictive value: proportion of negative classifications that are true negatives.
• Calibration: agreement between predicted probabilities and observed frequencies.
• Discrimination: ability to rank people with outcomes above those without outcomes.

Sensitivity is:

Specificity is:

These measures require a defined reference standard and a specified threshold. Predictive values depend strongly on outcome prevalence, so performance in one population may not transfer to another.

Validation of AI systems

Validation should proceed in stages:

1. Technical validation: Does the system operate as designed?
2. Analytical validation: Are outputs accurate against an appropriate reference standard?
3. External validation: Does performance hold in different sites, populations and time periods?
4. Clinical validation: Does use improve decisions or outcomes?
5. Implementation evaluation: Does it fit workflow without unacceptable burden or harm?
6. Post-deployment surveillance: Does performance drift as practice, population or data change?

A retrospective test on historical data is insufficient evidence of clinical benefit. Prospective evaluation should examine unintended consequences, subgroup performance, overrides, delays and equity. Model drift may result from changes in coding, equipment, prevalence, clinical practice or population characteristics.

9. Bias, fairness and equity

Algorithmic bias can arise from unrepresentative training data, measurement differences, historical discrimination, missing data, proxy variables, inappropriate labels or unequal access to follow-up. Removing a protected characteristic does not necessarily remove bias because other variables may act as proxies.

A fair evaluation asks:

• Who was included and excluded from development data?
• Is the reference standard equally valid across groups?
• Are errors distributed differently by race, sex, age, disability, language or socioeconomic status?
• Does the intervention triggered by the model have equal availability?
• Who benefits and who bears the risks?

Fairness is not a single technical formula. Some fairness criteria cannot all be satisfied simultaneously when groups have different base rates. Governance must therefore combine statistical assessment with ethical reasoning, patient involvement and legal requirements.

10. Privacy, cybersecurity and information governance

Privacy concerns appropriate control over personal information; confidentiality concerns preventing unauthorised disclosure; security concerns protecting confidentiality, integrity and availability. These objectives are connected but distinct.

Core controls include:

• data minimisation and purpose limitation;
• role-based and least-privilege access;
• strong authentication and secure credentials;
• encryption in transit and at rest;
• patching, backup and recovery;
• network segmentation;
• audit logs and anomaly detection;
• staff training and phishing resistance;
• tested incident-response plans;
• supplier and cloud-service assurance;
• secure disposal and retention controls.

Cybersecurity is a patient-safety responsibility. Ransomware, for example, can delay medication administration, diagnostic testing, surgery and emergency communication. A business-continuity plan should identify critical clinical functions, paper or alternative workflows, communication channels, recovery priorities and reconciliation procedures after restoration.

Organisations must comply with applicable privacy, health-record, medical-device, cybersecurity and AI regulations. Legal compliance is a baseline, not proof that a system is ethically or clinically safe.

11. Technology-related incident management

Digital incidents should be reported and investigated without automatically blaming individuals. A useful investigation examines the work system: interface design, training, staffing, alerts, policies, data quality, interoperability, timing and recovery arrangements.

Examples include selecting the wrong patient, accepting an incorrect default, acting on an outdated result, missing an alert, or relying on a system during an unrecognised outage. Near misses are valuable because they reveal weak controls before injury occurs.

Corrective actions should be specific and monitored. “Retrain staff” alone is usually weak because it relies on memory and does not correct design defects. Stronger controls include forcing functions, clearer patient identifiers, safer defaults, independent verification, improved interfaces and automated reconciliation.

Illustrative case: predictive deterioration alert

Background

A district hospital introduced an AI-based deterioration alert integrated into its EHR. The system generated a risk score every hour for adult medical inpatients and displayed an alert to the ward nurse when the score exceeded a threshold.

Problem

During the first month, nurses reported frequent alerts for patients who did not deteriorate. Some alerts appeared after a patient had already been reviewed, while patients with incomplete observations sometimes received no alert. Staff began dismissing notifications. A review also found that the model had been developed mainly using data from tertiary hospitals and performed less well in the district hospital’s older population.

Analysis

The quality team examined alert frequency, response times, false negatives, missing observations, subgroup performance and workflow. They found that the model treated missing observations as if they represented low risk, an unsafe assumption. The alert screen did not show which variables contributed to the score, and escalation responsibilities were unclear. The technical system was functioning, but the sociotechnical system was not safe.

Approach and solution

The hospital suspended autonomous escalation and retained the tool as a decision aid. It changed missing-data handling so that absent essential observations generated a data-quality task rather than a low-risk score. The alert threshold was reviewed using local data. Nurses received a concise explanation, and the response pathway specified assessment by a clinician within a defined period. The hospital established weekly monitoring by age, sex, ethnicity, ward and completeness of observations. A downtime procedure and manual observation checklist were added.

Outcome

After redesign, alert volume decreased, response documentation improved and staff reported greater trust. The hospital did not claim that the tool reduced mortality until a prospective evaluation compared outcomes with the pre-implementation period while monitoring balancing measures such as workload and delayed care.

Learning points

The case demonstrates that validation is local and continuous; missing data can be a safety signal; explainability supports appropriate use; and human oversight must be designed into workflow rather than added as a slogan.

Governance and implementation framework

A hospital introducing a digital quality intervention should establish a multidisciplinary governance group involving clinical staff, quality and safety leaders, information technology, cybersecurity, privacy, procurement, patients and affected communities. The group should approve purpose, evidence, risk classification, accountability, procurement conditions and monitoring requirements.

A practical implementation cycle is:

1. Define the clinical or operational problem before selecting technology.
2. Map the current workflow and identify hazards.
3. Specify data requirements, users and success measures.
4. Assess privacy, security, equity and regulatory obligations.
5. Test usability and interoperability with representative users.
6. Validate performance locally and define acceptable limits.
7. Pilot with clear stop criteria and incident reporting.
8. Train users and communicate limitations.
9. Monitor outcomes, workload, equity, drift and unintended effects.
10. Review, modify, retire or replace the system when evidence no longer supports it.

The system owner must be named. Responsibilities should cover data stewardship, model performance, vendor communication, clinical escalation, incident response and retirement. Procurement contracts should address data ownership, audit access, cybersecurity, service availability, model updates, change notification and exit arrangements.

Challenges and limitations

Digital transformation can increase rather than reduce workload, particularly when documentation requirements are poorly designed. Interoperability may expose contradictory records rather than create a single truth. Dashboards can encourage target gaming. AI may reproduce historical inequities or appear objective because its calculations are complex. Privacy protections can conflict with data availability, while excessive access can conflict with confidentiality.

No metric captures quality completely. Quantitative performance must be combined with qualitative inquiry, patient experience, professional judgement and safety reporting. Technology should be judged by its effect on care, not by adoption alone.

Contemporary developments

Current developments include patient-generated data from wearable devices, natural-language processing of clinical notes, remote monitoring, generative AI for drafting documentation, federated learning and digital twins. These developments create opportunities but also new concerns about hallucinated content, unclear provenance, model confidentiality, consent for secondary use and the environmental cost of computation.

Generative AI should not be treated as an authoritative clinical source. Outputs require verification against the patient record and trusted evidence. Sensitive information should not be entered into unapproved tools. Organisations need explicit policies for disclosure, review, authorship, retention and incident reporting.

Chapter Summary

Digital quality is the management of information technologies as sociotechnical systems that influence clinical decisions and patient outcomes. EHRs improve availability and continuity but introduce risks involving usability, data integrity, alert fatigue and downtime. CDS is safest when it provides relevant information to the right person at the right time and preserves professional judgement. Telehealth expands access but requires clinical triage, privacy, escalation and equity safeguards.

Interoperability requires technical, syntactic and semantic compatibility, together with patient matching and data provenance. Dashboards are useful only when indicators have clear definitions, denominators, targets, owners and action rules. Automation should include validation, exception handling, auditability and human escalation.

AI requires technical, analytical, external, clinical and implementation validation, followed by continuous monitoring for drift and subgroup inequity. Sensitivity, specificity, predictive value, calibration and discrimination must be interpreted in context. Privacy and cybersecurity are patient-safety obligations, and governance must address accountability, bias, transparency, consent and regulatory compliance. Safe digital transformation begins with a defined care problem and ends with evidence that the technology improves care without unacceptable new harms.

Key Terms

Electronic health record: A longitudinal digital record of a person’s health information.

Clinical decision support: Digital information or recommendations designed to assist healthcare decisions.

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

Semantic interoperability: Exchange of information with shared meaning.

Telehealth: Delivery of healthcare through telecommunications and digital technologies.

Digital dashboard: A visual display of selected indicators for monitoring and action.

Automation: Performance of tasks with limited human intervention.

Artificial intelligence: Computational methods that perform tasks associated with human cognitive functions.

Machine learning: Methods that learn patterns from data to make predictions or classifications.

Algorithmic bias: Systematic and unfair differences in algorithmic performance or effects.

Model drift: Change in model performance as data, populations or practice change.

Sensitivity: The proportion of true cases correctly identified.

Specificity: The proportion of non-cases correctly classified as negative.

Calibration: Agreement between predicted probabilities and observed frequencies.

Alert fatigue: Reduced attention caused by excessive or low-value alerts.

Data provenance: Information about the origin, history and transformations of data.

Role-based access: Restriction of information access according to a user’s authorised role.

Human oversight: Meaningful human ability and responsibility to review, challenge or override technology.

Sociotechnical system: A system formed by interacting technical, human, organisational and social elements.

Technology-related incident: A safety event or near miss involving a digital system or technology-supported process.

Review Questions

Short Answer Questions

1. Define semantic interoperability.
2. State two benefits and two risks of electronic health records.
3. What is alert fatigue?
4. Name four dimensions of data quality.
5. Why is post-deployment monitoring necessary for AI systems?

Long Answer Questions

1. Explain how a hospital should evaluate an electronic health record from technical, information, workflow and outcome perspectives.
2. Discuss the design principles of safe clinical decision support.
3. Explain the relationship between telehealth, access, clinical appropriateness and health equity.
4. Describe the stages of validating an AI system for clinical use.
5. Discuss how cybersecurity failures can become patient-safety incidents and identify appropriate controls.

Analytical / Application Questions

1. A dashboard reports a fall rate of 4 per 1,000 occupied bed-days, but the hospital has recently changed its denominator definition. Analyse the risks of comparing this rate with previous months and recommend corrective actions.
2. An AI readmission model performs well overall but has substantially lower sensitivity for patients who do not speak the dominant hospital language. Analyse possible causes and propose a fair validation and governance response.
3. During an EHR outage, clinicians use paper medication charts and later re-enter orders. Design a safety plan covering immediate operations, communication, reconciliation, incident reporting and improvement after recovery.

Multiple Choice Questions with Answers

Question 1

Which dimension of interoperability concerns shared meaning of exchanged data?

A. Technical interoperability

B. Semantic interoperability

C. Physical interoperability

D. Financial interoperability

Correct Answer: B. Semantic interoperability

Question 2

Which practice is most likely to reduce clinical alert fatigue?

A. Increasing the number of interruptive alerts

B. Displaying every alert to every staff member

C. Improving alert relevance and presenting it at the appropriate workflow point

D. Removing all clinical alerts

Correct Answer: C. Improving alert relevance and presenting it at the appropriate workflow point

Question 3

A rate of inpatient falls per 1,000 occupied bed-days uses which denominator?

A. Number of admissions

B. Number of discharges

C. Occupied bed-days

D. Number of hospital employees

Correct Answer: C. Occupied bed-days

Question 4

Sensitivity is calculated as:

A. True positives ÷ (true positives + false negatives)

B. True negatives ÷ (true negatives + false positives)

C. True positives ÷ (true positives + false positives)

D. False positives ÷ (true negatives + false positives)

Correct Answer: A. True positives ÷ (true positives + false negatives)

Question 5

Which statement about predictive values is correct?

A. They are independent of disease prevalence

B. They depend partly on outcome prevalence

C. They are identical to sensitivity in all settings

D. They do not require a classification threshold

Correct Answer: B. They depend partly on outcome prevalence

Question 6

Which is an example of semantic interoperability?

A. Two systems use network cables

B. Two systems can exchange a file

C. Two systems interpret a laboratory result with the same test meaning and units

D. Two systems have different patient identifiers

Correct Answer: C. Two systems interpret a laboratory result with the same test meaning and units

Question 7

What is the safest interpretation of an AI risk score?

A. It is an unavoidable clinical diagnosis

B. It is a probability or classification that requires contextual professional judgement

C. It replaces the need for patient assessment

D. It is valid indefinitely after initial testing

Correct Answer: B. It is a probability or classification that requires contextual professional judgement

Question 8

Which control best reflects the principle of least privilege?

A. Giving every employee access to all records

B. Allowing access only to information needed for an authorised role

C. Disabling audit logs

D. Sharing one account across a clinical team

Correct Answer: B. Allowing access only to information needed for an authorised role

Question 9

Why is external validation of an AI model important?

A. It guarantees that the model is unbiased

B. It tests whether performance transfers to different populations or settings

C. It eliminates the need for clinical evaluation

D. It prevents all future model drift

Correct Answer: B. It tests whether performance transfers to different populations or settings

Question 10

Which feature is essential to safe automation in healthcare?

A. No possibility of human intervention

B. Unrecorded system changes

C. Exception handling and a defined escalation route

D. Permanent use of the original software version

Correct Answer: C. Exception handling and a defined escalation route