Artificial Intelligence and Intelligent Hospital Systems

Introduction and Context

Artificial intelligence (AI) is increasingly embedded in hospital information systems (HIS), electronic health records (EHRs), diagnostic devices, scheduling platforms, revenue-cycle systems and patient-facing applications. In healthcare, however, AI is not simply a software upgrade. It changes how clinical evidence is produced, how work is allocated, how decisions are justified and how institutional risk is managed.

An intelligent hospital system combines data, computational models, workflow integration and human expertise to support decisions or perform tasks. Its value depends on more than predictive accuracy. A model may be technically accurate yet clinically unusable, operationally disruptive, inequitable or unsafe when deployed in a different hospital. Consequently, health-care managers must evaluate the complete sociotechnical system: data-generating processes, algorithms, users, workflows, governance, incentives and outcomes.

The most important contemporary applications include machine learning for prediction and risk stratification, natural-language processing (NLP) for clinical text, computer vision for images and video, automation for repetitive administrative and clinical processes, and generative AI for drafting, summarization and conversational assistance. Each offers benefits, but each introduces distinctive hazards involving validation, explainability, bias, privacy, cybersecurity, accountability and human oversight.

Learning Objectives

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

1. Distinguish major forms of AI used in hospitals and explain their data and workflow requirements.
2. Critically assess AI applications in diagnosis, documentation, forecasting and operations.
3. Select appropriate validation, performance and monitoring methods.
4. Explain distribution shift, bias, calibration, interpretability and automation risk.
5. Design governance arrangements that preserve human accountability and patient safety.
6. Evaluate a proposed AI system from clinical, managerial, ethical, legal and economic perspectives.

1. From Digital Hospitals to Intelligent Hospitals

A conventional HIS records transactions: registration, orders, results, medications, diagnoses, billing and discharge. An intelligent HIS additionally detects patterns, predicts events, recommends actions or generates content. It may answer questions such as:

• Which admitted patients are at elevated risk of deterioration?
• How many operating-room cases are likely to be delayed tomorrow?
• Which clinical notes contain evidence of a medication discrepancy?
• Can a radiological image be prioritized for specialist review?
• Can a draft discharge summary be prepared from verified record data?

The distinction between automation and augmentation is important. Automation performs a task with limited human intervention; augmentation assists a professional who remains responsible for interpretation and action. A system that automatically routes a suspected stroke image for urgent review is different from one that independently declares a diagnosis and initiates treatment.

An intelligent system can be represented conceptually as a feedback loop:

Data → model → output → human or organizational action → outcome data → evaluation and updating

The loop matters because deployment changes behaviour. Clinicians may order fewer tests when a prediction tool is trusted, or more tests when it is distrusted. Documentation practices may change after an NLP system is introduced. Thus, model performance measured before deployment may not equal performance after deployment.

Important terminology

• Artificial intelligence: computational methods that perform tasks associated with human intelligence, such as prediction, perception, language processing or planning.
• Machine learning: methods that infer patterns from data rather than relying exclusively on hand-coded rules.
• Supervised learning: learning from examples paired with target labels.
• Unsupervised learning: discovering structure without a specified target label.
• Foundation model: a large model trained on broad data and adapted to multiple downstream tasks.
• Inference: producing an output from a trained model and new input data.
• Model drift: deterioration or change in model behaviour as data, populations or practices evolve.
• Clinical decision support: information or recommendations intended to assist health-care decisions.
• Human-in-the-loop: a design in which a person reviews, interprets or authorizes an AI output.

2. Core AI Modalities in Hospitals

2.1 Machine learning

Machine learning can support classification, regression, ranking, clustering and anomaly detection. A classification model may estimate whether a patient is at high risk of readmission; a regression model may estimate length of stay; clustering may identify utilization patterns; anomaly detection may flag an unusual laboratory result or billing transaction.

A supervised model is trained using input variables, often called features, and a target outcome. For patient i, the model produces a prediction \hat{y}_i = f(x_i), where x_i is the feature vector, f is the learned function and \hat{y}_i is the predicted outcome. The prediction is not the outcome itself and should not be interpreted as certainty.

For binary risk prediction, a common performance measure is the area under the receiver operating characteristic curve (AUROC). Sensitivity and specificity are defined as:

Here, TP is true positives, FN false negatives, TN true negatives and FP false positives. Sensitivity measures the proportion of actual positive cases detected; specificity measures the proportion of actual negative cases correctly rejected. These metrics are conditional on a chosen threshold. They do not by themselves establish clinical usefulness.

When predicted probabilities are used, calibration is essential. A well-calibrated group of patients predicted to have 20% risk should experience the event approximately 20% of the time, subject to sampling variation. A model can discriminate well but be poorly calibrated, especially when transported to another hospital.

For imbalanced outcomes, precision and recall can be more informative than accuracy. Precision is:

Accuracy can be misleading when the event is rare. For example, a model that labels every patient “no sepsis” may appear accurate in a population where sepsis is uncommon while failing every patient with sepsis.

2.2 Natural-language processing

NLP converts unstructured language into searchable, structured or generative outputs. Clinical text includes progress notes, discharge summaries, radiology reports, pathology reports, referral letters, messages and dictated speech. NLP applications include entity recognition, medication extraction, symptom coding, information retrieval, cohort identification, summarization and documentation assistance.

Clinical language is difficult because of negation, uncertainty, abbreviations, copied-forward text, temporal references and local terminology. “No evidence of pneumonia” should not be coded as pneumonia; “history of diabetes” is not equivalent to active uncontrolled diabetes; and “rule out pulmonary embolism” is not a confirmed diagnosis.

A useful evaluation distinguishes extraction accuracy from clinical usefulness. A system may correctly identify a diagnosis mention yet fail to identify whether it is current, historical, suspected or negated. For generative summaries, factuality, omission, attribution and readability must all be assessed. The risk of hallucination—fluent but unsupported content—is particularly serious in clinical documentation.

2.3 Computer vision

Computer vision analyses medical images, pathology slides, photographs, video or sensor feeds. It can assist with image prioritization, segmentation, measurement, quality control and detection of abnormalities. A segmentation system identifies pixels or regions; a classification system assigns a category; a detection system identifies both category and location.

Image performance is sensitive to scanner type, acquisition protocol, demographic composition, disease prevalence and labeling practice. A model trained on one institution’s radiographs may learn markers of hospital or equipment rather than pathology. External validation using independent hospitals and time periods is therefore important.

2.4 Automation and robotic process automation

Rule-based automation and robotic process automation (RPA) can move data between systems, verify eligibility, route referrals, schedule appointments, reconcile lists and generate alerts. Unlike machine learning, RPA often follows explicit rules and is most suitable where processes are stable, inputs are structured and exceptions are well understood.

Automation should not be introduced merely because a task is repetitive. Process defects can be amplified at scale. Before automation, managers should map the process, remove unnecessary steps, define exception handling and establish rollback procedures. The design should record what the automation did, when it did it and which source data it used.

2.5 Generative AI and large language models

Generative AI produces text, code, images or other content from prompts and context. In hospitals, relatively lower-risk uses may include drafting patient education in approved language, summarizing a clinician’s own verified notes, creating meeting minutes or assisting information retrieval over governed institutional documents.

Higher-risk uses include autonomous diagnosis, medication recommendations, unreviewed discharge instructions, direct patient triage and generation of legal or billing assertions. Large language models may produce plausible errors, reveal sensitive information through poor configuration, reproduce bias or cite sources that do not exist. Retrieval-augmented generation can constrain answers to approved documents, but it does not eliminate errors: the retrieval set may be incomplete, outdated or misinterpreted.

3. Applications Across the Hospital Value Chain

Diagnosis and clinical decision support

AI can support diagnosis by detecting imaging patterns, ranking differential diagnoses, identifying deteriorating patients and matching patients to clinical pathways. The safest design usually places AI in a clearly defined role: prioritization, second reading, information retrieval or risk estimation. The system should state the intended population, target condition, time horizon and action associated with an alert.

The distinction between prediction and causation must be maintained. A model predicting readmission does not show that changing the predicted feature will reduce readmission. Likewise, a risk score should not automatically determine eligibility for a scarce service unless its consequences and fairness have been evaluated.

Documentation and coding

Ambient speech tools can transcribe encounters and propose notes. NLP can suggest codes or identify missing documentation. Benefits may include reduced clerical burden and improved information availability. Risks include incorrect attribution, missing negation, overdocumentation, privacy loss and the propagation of errors into the legal medical record.

A robust workflow labels generated text as draft, displays source evidence where feasible, requires clinician review and records who accepted or edited the content. Copy-forward behaviour should be monitored because generated errors can become durable facts.

Forecasting and capacity planning

Forecasting models estimate emergency arrivals, bed demand, operating-room utilization, staffing requirements, supply consumption and discharge volume. A basic forecasting error measure is mean absolute error:

where n is the number of forecast periods, y_t is the observed value in period t, and \hat{y}_t is the forecast. MAE has the same units as the outcome—for example, patients or occupied beds—making it operationally interpretable. It penalizes errors linearly and is less sensitive to extreme errors than mean squared error.

Forecasts should be evaluated against meaningful baselines, such as the previous week, seasonal average or established planning method. A more complex model is not justified unless it improves decisions, not merely a statistical score.

Operations and patient flow

AI may assist appointment scheduling, theatre sequencing, discharge prediction, ambulance diversion, inventory management and queue prioritization. Operational systems must be evaluated for second-order effects. Prioritizing one queue can lengthen another; optimizing bed occupancy may increase transfers or reduce surge capacity; minimizing staff cost may increase burnout and safety risk.

A useful managerial question is: “What objective is being optimized, for whom, over what period, and at whose expense?” Multi-objective settings should explicitly consider safety, equity, quality, staff workload, patient experience and financial sustainability.

4. Validation: From Technical Performance to Clinical Value

Validation should occur in stages.

Random splitting of records can produce optimistic results when the same patient appears in multiple records or when data from the future leak into training. Data leakage occurs when information unavailable at the prediction time influences the model. For example, a discharge diagnosis should not be used to predict deterioration earlier in the admission.

Validation must specify the unit of analysis, prediction time, outcome definition, missing-data strategy, threshold, follow-up period and intended action. Confidence intervals should accompany estimates where feasible. Subgroup performance should be examined by clinically relevant factors such as age, sex, ethnicity, language, disability, socioeconomic context, comorbidity and site—not to essentialize groups, but to detect unequal performance and access.

A model’s value can be understood through a decision framework. Let the action threshold reflect the relative consequences of false positives and false negatives. The threshold should be chosen with clinical stakeholders, not selected solely to maximize a generic metric. Decision-curve analysis or net-benefit approaches can help assess whether using a model improves decisions across plausible threshold probabilities, but such methods depend on assumptions about harms, benefits and patient preferences.

5. Explainability, Bias and Human Oversight

Explainability

Interpretability refers to how readily a person can understand a model’s operation; explainability refers to methods used to communicate why a particular output was produced. A simple rule or linear model may be intrinsically interpretable, whereas a complex neural network may require post hoc explanations.

Explanations can be global, describing general model behaviour, or local, describing one prediction. Feature importance does not prove causation. A plausible explanation can still be incomplete or unstable. In clinical settings, the most useful explanation may include the prediction, confidence or uncertainty, key contributing evidence, relevant time window, known limitations and recommended human checks.

Bias and fairness

Bias can enter through historical inequity, underrepresentation, measurement error, label choices, missingness, selective access to care and deployment decisions. If historical treatment reflects unequal access, learning to reproduce treatment may encode inequity rather than clinical need.

Fairness is not a single mathematical property. Possible criteria include comparable sensitivity across groups, comparable calibration, comparable false-positive rates or comparable benefit from intervention. These criteria can conflict, particularly when event prevalence differs. Therefore, fairness decisions require clinical, ethical and community input rather than a purely technical selection of one metric.

Mitigation options include better data collection, redefinition of labels, subgroup-specific evaluation, threshold review, reweighting, model revision, human review and restricting use. Removing a protected attribute does not remove bias because proxy variables may remain.

Human oversight

Human oversight must be meaningful, not ceremonial. A clinician who receives an unexplained alert at the busiest point in a shift may be unable to review it properly. Effective oversight requires adequate time, training, authority to override, access to underlying evidence and protection from inappropriate automation pressure.

Two risks are automation bias, excessive acceptance of machine output, and automation complacency, reduced vigilance because the system is assumed to be reliable. Interfaces should make uncertainty visible, distinguish recommendation from decision and support disagreement. Override reasons can be collected for learning, but a high override rate should not automatically be treated as clinician failure; it may indicate poor model fit.

6. AI Governance and Accountability

AI governance is the institutional system for deciding which AI applications may be acquired, developed, deployed, monitored, modified or retired. It should connect clinical governance, information governance, cybersecurity, procurement, ethics, legal review, quality improvement and patient safety.

A practical governance lifecycle includes:

1. Use-case definition: identify the problem, affected people, intended benefit, alternatives and unacceptable uses.
2. Risk classification: assess clinical severity, autonomy, scale, data sensitivity and reversibility of harm.
3. Data governance: establish lawful access, provenance, quality, retention, consent expectations and information security.
4. Technical and clinical evaluation: validate performance, calibration, subgroup behaviour, usability and workflow fit.
5. Approval and contracting: define responsibilities, audit rights, change notification, incident response and data-use restrictions.
6. Controlled deployment: begin with a limited population or silent mode where appropriate.
7. Monitoring: track performance, drift, alert burden, overrides, adverse events, complaints and equity indicators.
8. Review and retirement: reassess when the population, workflow, vendor or model changes.

A model card or system dossier should document intended use, contraindications, training data characteristics, evaluation settings, known limitations, thresholds, update history and contact responsibilities. A data protection impact assessment is appropriate when processing presents significant privacy risk. Cybersecurity controls should address prompt injection, data exfiltration, unauthorized access, model tampering, insecure interfaces and supply-chain vulnerabilities.

Responsibility cannot be transferred to an algorithm or vendor. The hospital remains accountable for the way it deploys a system, while clinicians remain responsible for professional decisions within their legal and ethical duties. Contracts should clarify liability, incident reporting, audit access, intellectual property, data ownership and service continuity.

7. Implementation and Change Management

Many AI projects fail because they begin with a technology rather than a defined problem. A disciplined implementation asks:

• What measurable problem exists?
• Which current workflow produces the problem?
• What is the baseline performance and cost?
• Who will use the output, and at what point?
• What action follows a positive or negative result?
• What happens when the system is unavailable?
• How will patients and staff be informed?

Process mapping and co-design with clinicians, administrators, patients, information specialists and safety personnel are essential. Usability testing should include realistic workload, interruptions and edge cases. Training should explain not only how to use a tool, but when not to trust it.

Economic evaluation should include licensing, integration, validation, training, monitoring, cybersecurity, workflow redesign and opportunity costs. Benefits may include time saved, improved throughput, avoided harm or better experience. A reduction in documentation time is not a benefit if it is offset by increased verification work or inaccurate records.

8. Illustrative Case: An AI Deterioration Alert

Background

A fictional 450-bed hospital proposes a machine-learning model that predicts deterioration within the next 12 hours using vital signs, laboratory results, nursing observations and recent orders. The objective is to help a rapid-response team review high-risk patients.

Problem

The retrospective model reports an AUROC of 0.86. However, the development data came from the same hospital over five years, and the outcome label is transfer to intensive care. Leaders are concerned about alert volume, unequal performance across wards and whether the model will reduce harm.

Analysis

The outcome “ICU transfer” is not identical to physiological deterioration; it also reflects bed availability, clinician preference and local escalation practices. The model may therefore learn organizational behaviour. The team discovers that some laboratory results entered after escalation were accidentally included, creating temporal leakage. The model also performs less well in a ward serving many older patients with incomplete observations.

The intended action is not automatic transfer. It is a structured review by a trained nurse, followed by escalation according to existing clinical protocols. This distinction reduces, but does not remove, risk.

Approach and solution

The hospital removes post-outcome variables, defines the prediction time precisely and conducts temporal and external validation. It assesses sensitivity, positive predictive value, calibration and alert burden by ward and patient subgroup. A silent prospective phase compares predictions with outcomes without notifying clinicians. A pilot then introduces alerts only during staffed hours, with a backup process for downtime.

The interface displays the prediction time, major contributing data, missingness, uncertainty and a link to the relevant chart evidence. The nurse can acknowledge, defer or override the alert and record a reason. The governance committee sets a maximum acceptable alert burden and requires weekly safety review during the pilot.

Outcome

The pilot identifies that many alerts occur after a clinician has already recognized the problem. Rather than treating this as a model success, the hospital redesigns the alert to suppress redundant notifications and focuses on patients not already under active escalation. The system improves response time in a selected ward, but the hospital does not claim reduced mortality because the pilot was not designed or powered to establish that outcome.

Learning points

1. A high AUROC does not establish clinical benefit.
2. The label must represent the decision problem, not merely an available administrative event.
3. Temporal leakage can invalidate apparently strong performance.
4. Alert burden and workflow fit are safety outcomes.
5. Human review, subgroup analysis and monitoring are design requirements, not optional additions.

9. Contemporary Developments and Research Directions

Foundation models are expanding the scope of clinical language and multimodal systems. Their potential includes cross-document reasoning, structured extraction, image-text analysis and personalized education. Yet scale does not guarantee reliability. Research priorities include reliable uncertainty estimation, provenance-linked generation, privacy-preserving learning, federated evaluation, causal decision support, robust subgroup testing and methods for measuring patient-centred outcomes.

Digital twins and hospital simulation may combine real-time data with operational models to test bed, staffing or pathway decisions before implementation. Federated learning can allow institutions to train collaboratively without centralizing all patient data, although it introduces challenges involving governance, heterogeneity, communication, security and verification of local performance.

Regulatory approaches increasingly emphasize risk, intended use, lifecycle monitoring and transparency. Health-care managers should monitor applicable national rules, professional standards, medical-device requirements, privacy legislation and procurement guidance. Compliance is necessary but not sufficient: an approved system can still be poorly integrated or inappropriate for a particular local use.

Best-Practice Checklist

Before deployment, an organization should be able to answer “yes” to the following:

• Is the clinical or operational problem clearly defined?
• Is AI preferable to a simpler intervention or process redesign?
• Are training, validation and outcome data appropriate and traceable?
• Has temporal leakage been excluded?
• Has performance been tested externally and across relevant subgroups?
• Are calibration, uncertainty and clinically meaningful thresholds understood?
• Is the human role, authority and escalation pathway explicit?
• Are privacy, cybersecurity, procurement and liability addressed?
• Is generated content clearly marked and reviewed before entering the record?
• Are monitoring, incident response, rollback and retirement plans funded?
• Are patients and staff given appropriate information about material AI use?
• Has the organization measured real-world benefit rather than model accuracy alone?

Chapter Summary

Artificial intelligence can make hospital information systems more predictive, perceptive, efficient and responsive. Machine learning supports risk estimation and forecasting; NLP supports extraction, retrieval and documentation; computer vision supports image and video analysis; automation streamlines defined workflows; and generative AI assists with language and knowledge tasks. These technologies differ in data requirements, failure modes and appropriate levels of autonomy.

Safe adoption requires a sociotechnical perspective. Validation must progress from internal testing to temporal, external, prospective and impact evaluation. Performance should include discrimination, calibration, subgroup behaviour, workload and patient outcomes. Explainability must support informed review rather than create a false impression of certainty. Bias may arise from data, labels, access and institutional practice, and fairness requires explicit value judgments.

Human oversight must be empowered and meaningful. Governance should cover use-case definition, data protection, cybersecurity, procurement, accountability, monitoring, incident response and retirement. The central managerial test is not whether a model is impressive, but whether its use improves care or operations without unacceptable loss of safety, equity, privacy, professional judgment or public trust.

Key Terms

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

Machine learning: Methods that learn patterns from data to produce predictions or decisions.

Natural-language processing: Computational analysis and generation of human language.

Computer vision: Computational analysis of images, video or visual data.

Generative AI: AI that creates new text, images, code or other content.

Foundation model: A broadly trained model adapted to multiple tasks.

Clinical decision support: Information or recommendations intended to assist care decisions.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Sensitivity: The proportion of actual positive cases correctly identified.

Specificity: The proportion of actual negative cases correctly identified.

Data leakage: Use of information in model development that would not be available at prediction time.

Distribution shift: Change between development data and deployment data.

Automation bias: Excessive reliance on an automated recommendation.

Explainability: Communication of reasons or evidence associated with a model output.

Model drift: Change or deterioration in model behaviour over time.

Human-in-the-loop: Design requiring meaningful human review or authorization.

Algorithmic fairness: Evaluation and management of unequal performance or impact across groups.

AI governance: Institutional oversight of AI throughout its lifecycle.

Hallucination: Unsupported or fabricated output presented by a generative system.

Retrieval-augmented generation: Generation grounded in retrieved documents or knowledge sources.

Review Questions

Short Answer Questions

1. Distinguish automation from augmentation in a hospital AI system.
2. Why can accuracy be misleading for a rare clinical event?
3. What is data leakage, and why does it threaten validation?
4. Explain the difference between discrimination and calibration.
5. Identify three governance activities required after AI deployment.

Long Answer Questions

1. Critically compare machine learning, NLP, computer vision and generative AI in hospital applications.
2. Explain how an AI documentation assistant should be validated and integrated into the medical-record workflow.
3. Discuss the sources of bias in a clinical prediction model and evaluate possible mitigation strategies.
4. Design a lifecycle governance framework for an AI system used in patient-flow management.
5. Explain why technical performance does not necessarily demonstrate clinical or organizational value.

Analytical / Application Questions

1. A deterioration model has high AUROC but generates 80 alerts per ward per shift, most of which are clinically redundant. Analyse the problem and propose a safer deployment strategy.
2. A generative AI tool drafts discharge summaries, but an audit finds occasional medication omissions and invented follow-up appointments. Decide whether deployment should continue and specify controls required.
3. A hospital’s readmission model performs better for privately insured patients than for publicly insured patients. Develop an evaluation and governance response that addresses performance, fairness and access without simply removing the insurance variable.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes calibration from discrimination in a clinical prediction model?

A. Calibration concerns probability accuracy, while discrimination concerns ranking cases by risk.

B. Calibration measures privacy protection, while discrimination measures cybersecurity.

C. Calibration applies only to images, while discrimination applies only to text.

D. Calibration is the same as sensitivity at every threshold.

Correct Answer: A. Calibration concerns probability accuracy, while discrimination concerns ranking cases by risk.

Question 2

Which situation is an example of data leakage?

A. Using age recorded at admission to predict deterioration.

B. Using a discharge diagnosis to predict deterioration earlier in the admission.

C. Using vital signs available at the prediction time.

D. Testing a model on a later time period.

Correct Answer: B. Using a discharge diagnosis to predict deterioration earlier in the admission.

Question 3

Why can accuracy be misleading for a rare event such as sepsis?

A. Accuracy cannot be calculated for binary outcomes.

B. A model can achieve high accuracy by predicting the common negative class for nearly everyone.

C. Accuracy always equals sensitivity.

D. Rare events have no measurable outcomes.

Correct Answer: B. A model can achieve high accuracy by predicting the common negative class for nearly everyone.

Question 4

Which is the most appropriate use of a generative AI draft in a medical record?

A. Automatically signing it without review.

B. Entering it as verified fact regardless of source evidence.

C. Clearly marking it as a draft and requiring clinician verification.

D. Allowing it to replace all source documentation.

Correct Answer: C. Clearly marking it as a draft and requiring clinician verification.

Question 5

What is a principal purpose of external validation?

A. To increase the number of model parameters.

B. To test transportability to an independent setting or population.

C. To eliminate the need for monitoring.

D. To guarantee causal effectiveness.

Correct Answer: B. To test transportability to an independent setting or population.

Question 6

Which metric is expressed in the same units as the forecasted outcome?

A. Mean absolute error.

B. AUROC.

C. Sensitivity.

D. Specificity.

Correct Answer: A. Mean absolute error.

Question 7

Which design best represents meaningful human oversight?

A. A clinician must accept every alert with no override option.

B. The system hides uncertainty to prevent disagreement.

C. A trained professional reviews evidence, can override the output and follows an escalation pathway.

D. A vendor alone decides whether an alert is clinically appropriate.

Correct Answer: C. A trained professional reviews evidence, can override the output and follows an escalation pathway.

Question 8

Why might removing ethnicity from a model fail to remove bias?

A. Models cannot operate without ethnicity.

B. Proxy variables and unequal data-generating processes may still encode related information.

C. Removing variables always increases discrimination.

D. Fairness is unrelated to input data.

Correct Answer: B. Proxy variables and unequal data-generating processes may still encode related information.

Question 9

Which activity belongs most clearly to post-deployment AI governance?

A. Monitoring drift, alert burden, subgroup performance and incidents.

B. Selecting a random font for the interface.

C. Deleting all audit logs.

D. Preventing users from reporting errors.

Correct Answer: A. Monitoring drift, alert burden, subgroup performance and incidents.

Question 10

A model produces TP = 80 and FN = 20. What is its sensitivity?

A. 20%.

B. 50%.

C. 80%.

D. 100%.

Correct Answer: C. 80%.