Infectious Diseases and Antimicrobial Stewardship

Infectious diseases combine microbiology, genomic sequencing, antimicrobial susceptibility testing, clinical context, and population surveillance. Machine-learning models can estimate resistance or prescribing risk, while conventional rules and automated pipelines can detect bug-drug mismatches or construct phylogenies. Automation, statistical modeling, and clinical AI are not interchangeable categories, and each requires evidence matched to its actual function. Short-horizon forecasting and nowcasting can inform operations, but no model reliably predicts the timing and location of novel outbreak emergence.

Learning Objectives

After reading this chapter, readers will be able to:

  • Evaluate AI systems for antimicrobial stewardship and antibiotic selection
  • Understand outbreak prediction and genomic surveillance AI applications
  • Assess rapid pathogen identification and resistance prediction tools
  • Navigate AI applications in infection control and hospital epidemiology
  • Recognize the role of AI in pandemic preparedness and response
  • Apply evidence-based frameworks for infectious diseases AI adoption

The Clinical Context: Infectious diseases combine clinical complexity (host factors, pathogen factors, antimicrobial pharmacology) with rich data sources (microbiology, genomics, epidemiology). AI applications span individual patient care to population-level outbreak response.

What Works Well:

Application Evidence Level Key Benefit
AI-supported stewardship Variable Can improve selected prediction or prescribing-process measures
Genomic surveillance platforms Strong computational infrastructure Reproducible tracking of pathogen evolution, not necessarily AI
MALDI-TOF identification Established laboratory automation Rapid species-level identification, distinct from emerging ML resistance prediction

What’s Emerging:

Application Status Notes
Genotype-to-phenotype resistance prediction Research Promising but inconsistent clinical validation
MALDI-TOF resistance prediction Research Site and temporal drift require local monitoring
Outbreak forecasting Variable Performance depends on target, horizon, place, and reporting conditions
HAI risk prediction Moderate Implementation studies ongoing
Imaging-based parasitic-disease staging Retrospective multicenter + reader study WHO–IWGE CE1–CE5 on non-contrast CT; not a treatment trial (Wang, Z. et al., 2026)

The Bottom Line: AI-supported stewardship can improve selected prediction and prescribing-process measures, while genomic surveillance and MALDI-TOF identification also rely on important non-AI computational infrastructure. Resistance prediction from sequencing or MALDI-TOF spectra requires local and temporal validation and does not replace reference AST. Forecasting claims must specify the target and horizon, and alert burden should be measured in the actual workflow.

Introduction

An estimated 1.27 million deaths in 2019 were directly attributable to bacterial antimicrobial resistance (Murray et al., 2022). Stewardship programs must translate local microbiology and patient context into timely prescribing decisions, while genomic surveillance supports pathogen tracking at population scale. The current landscape is defined less by a single AI capability than by a stack of laboratory automation, rules-based CDS, statistical prediction, machine learning, and expert response. Clinical benefit depends on how that stack changes a decision, not on whether one component is labeled AI.

Part 1: Antimicrobial Stewardship AI

The Stewardship Imperative

The 2019 burden estimate above concerns deaths attributable to bacterial antimicrobial resistance, not a measured annual effect of AI or stewardship. The CDC Core Elements of Hospital Antibiotic Stewardship Programs defines program infrastructure and actions, but it does not support one universal percentage reduction for an AI tool. AI can augment stewardship by processing microbiology and clinical data at scale, provided that the resulting recommendation is locally validated and connected to accountable review.

Decision Support Applications

Current AI stewardship tools:

  1. Bug-drug mismatch detection
    • Real-time alerts when prescribed antibiotic lacks coverage for cultured organism
    • Integration with lab information systems
    • Performance: High sensitivity but alert fatigue with low positive predictive value
  2. Duration of therapy monitoring
    • Flags antibiotic courses exceeding guideline recommendations
    • Supports de-escalation and IV-to-oral conversion
  3. Empiric therapy recommendations
    • Antibiotic selection based on local antibiogram, patient factors, indication
    • Updates dynamically as resistance patterns shift

IDSA/SHEA Stewardship Guidelines

IDSA/SHEA Guidelines on Antimicrobial Stewardship

The joint IDSA/SHEA guidelines on antimicrobial stewardship programs (Barlam et al., 2016) provide the framework for AI implementation:

Core recommendations:

  • Preauthorization and prospective audit with feedback remain gold standards
  • Clinical decision support should be integrated into EHR workflows
  • Local antibiograms should drive empiric therapy recommendations
  • Stewardship interventions require ID or pharmacy oversight

Implications for AI:

  • AI tools should support, not replace, prospective audit processes
  • Automated recommendations require human stewardship review
  • Local validation essential: national models may not reflect institutional resistance patterns
  • Alert frequency must balance sensitivity with actionability

Related 2024 guidance: SHEA published a pandemic-preparedness position statement addressing data infrastructure and surveillance systems (SHEA, 2024). This is related policy guidance, not an update to the 2016 stewardship guideline.

Implementation Evidence

Study Intervention Outcome
Timbrook et al., 2017 Non-AI comparator: rapid molecular diagnostic plus stewardship program Lower mortality in the coupled diagnostic-response evidence base
Corbin et al., 2022 ML personalized antibiograms Improved empiric coverage prediction

The Timbrook meta-analysis demonstrates the value of coupling a faster diagnostic result to an antimicrobial-stewardship response pathway. It is not evidence that an AI intervention caused the mortality difference. Diagnostic speed creates value only when a governed workflow converts the result into an appropriate treatment decision.

A 2025 systematic review of AI-driven antimicrobial stewardship applications found empirical antibiotic selection, resistance prediction, dose optimization, and vancomycin dosing support as the most common use cases. The review identified implementation studies showing potential for improved prescribing appropriateness, but most evidence remained observational and dependent on local data quality, resistance ecology, and stewardship oversight (Harandi et al., 2025).

A separate 2025 systematic review and meta-analysis included 80 observational, cohort, or retrospective studies and pooled model discrimination and diagnostic-accuracy measures across heterogeneous stewardship tasks. Because it did not establish randomized patient-outcome benefit, its performance estimates should not be converted into a claim that AI reduces resistance, mortality, or C. difficile infection (Pennisi et al., 2025).

A prospective longitudinal observational implementation in a nationwide Israeli outpatient network provides workflow evidence beyond retrospective accuracy. Among 19,287 cultured UTI cases, prescriptions that matched UTI Smart-Set (UTIS) recommendations were associated with fewer bug-drug mismatches; recommendation-concordant prescribing in the broader cohort was also associated with less ciprofloxacin use. Clinicians chose whether to follow the tool, so selection and confounding remain possible; the study did not assess clinical cure, recurrent culture, or antibiotic switching (Shapiro Ben David et al., 2025). UTIS supports locally governed empiric prescribing, but it does not establish patient-outcome benefit or transportability beyond the health system in which it was trained and deployed.

Clinical interpretation: AI stewardship tools should be treated like an extension of prospective audit and feedback, not a replacement for ID physician or pharmacist review. National models will not capture local antibiograms, outbreak patterns, or formulary constraints unless they are locally adapted and monitored.

Human-factors evidence is similarly bounded. In a randomized study using 12 simulated intravenous-to-oral switch vignettes, clinicians could identify and ignore incorrect advice, and adoption depended on clinical evidence and usability rather than explanation displays alone. The experiment did not measure prescribing behavior or patient outcomes in clinical care (Bolton et al., 2025).

Alert Fatigue in Stewardship

The fundamental challenge is that stewardship alerts compete with other clinical alerts. A review of 17 studies reported override rates of 49%–96% for drug-safety alerts, not specifically for AI antimicrobial recommendations (van der Sijs et al., 2006). Even in a primary-care system deliberately tuned to suppress false positives, 60% of serious drug-drug interaction alerts were overridden, and roughly one third of those overrides were judged inappropriate (Slight et al., 2013). Local acceptance, action, and error rates matter more than a transferred alert-fatigue statistic.

Mitigation strategies:

  • Tiered alerting (hard stops for critical mismatches, soft alerts for duration)
  • Pharmacist review before alert delivery
  • Bundled recommendations rather than individual alerts

Part 2: Genomic Surveillance and Phylogenetic Analysis

The Genomic Surveillance Revolution

COVID-19 transformed pathogen genomics from research tool to clinical necessity. The infrastructure built for SARS-CoV-2 surveillance now applies to other pathogens.

Key Platforms

Nextstrain:

The Nextstrain platform (Hadfield et al., 2018) provides real-time genomic epidemiology through:

  • Automated phylogenetic tree construction
  • Geographic and temporal visualization
  • Clade and variant assignment
  • Open-source analysis pipelines

Nextstrain supports maintained pathogen builds and community analyses across multiple organisms. Throughput changes by build, infrastructure, and data availability, so a fixed daily sequence count is not a stable measure of capability.

GISAID:

GISAID is an important repository for influenza and SARS-CoV-2 sequence data. Repository totals change continuously and should be read from the source at the time of analysis. Downstream systems can assign:

  • Pango lineages
  • WHO variant designations
  • Clade membership

Africa CDC Pathogen Genomics Initiative:

Regional surveillance demonstrates computational genomics at scale. Nextstrain Africa CDC builds track pathogen evolution across the continent with automated quality control and lineage assignment (Wilkinson et al., 2021). The platform’s automation should not be treated as evidence of a machine-learning intervention unless a specific learned component is evaluated.

Automated Phylogenetic Analysis

Automated computation and, in some components, machine learning can accelerate genomic-surveillance workflows:

  1. Sequence quality control
    • Automated detection of frameshifts, premature stops, contamination
    • Flagging for manual review
  2. Lineage assignment
    • Pangolin (SARS-CoV-2), Nextclade, custom classifiers
    • High-throughput assignment followed by quality control and expert curation
  3. Transmission cluster detection
    • Genetic-distance analysis to support, not define by itself, outbreak hypotheses
    • Integration with epidemiologic data
  4. Recombination detection
    • Identification of mosaic genomes
    • Critical for SARS-CoV-2 and influenza

Resistance Gene Detection

Whole-genome sequencing enables genotypic resistance prediction:

Examples of molecular and genomic resistance signals:

Pathogen or group Signal Evidence boundary
M. tuberculosis Catalogued variants associated with drug resistance Some molecular tests are clinically established; conventional gene detection is not necessarily AI
S. aureus mecA or mecC for methicillin resistance Gene detection supports a defined resistance mechanism, not all phenotypic resistance
Enterobacterales Carbapenemase genes Molecular detection is clinically used in defined workflows; expression and other mechanisms still matter
N. gonorrhoeae Multilocus genomic predictors Clinical implementation remains limited and requires current phenotype-linked validation

Limitations:

  • Novel resistance mechanisms not in databases
  • Phenotype-genotype discordance (expression levels, epistasis)
  • Turnaround time and incremental value depend on specimen processing, sequencing access, analysis, and the comparator laboratory workflow

Machine Learning for Resistance Prediction: Validation Challenges

A 2026 Lancet Infectious Diseases Series assessed artificial intelligence in infectious-disease diagnostics across pathogens and settings (Miglietta et al., 2026). The review describes substantial gaps between research promise and clinical readiness.

Current validation status by pathogen:

Pathogen ML Approach Validation Status Clinical Use
M. tuberculosis WGS-based DST Curated catalogues support specified variant-drug interpretations Used in some reference-laboratory pathways
E. coli Genotypic or clinical prediction Geographic and temporal transportability must be demonstrated Research and implementation evaluation
K. pneumoniae Genomic, spectral, or clinical prediction Diverse mechanisms and high study risk of bias limit transfer Predominantly retrospective research
S. aureus Defined gene detection or broader prediction mecA/mecC detection differs from learned prediction of other resistance Task-specific clinical or research use
P. aeruginosa Multidrug-resistance prediction Multiple intrinsic and acquired mechanisms complicate generalization Research evaluation

Validation remains inconsistent across settings and populations. Key barriers include:

  • Geographic diversity: Models trained in one resistance ecology can lose calibration or discrimination in another
  • Breakpoint variability: CLSI, EUCAST, and regional standards define resistance differently
  • Expression-level effects: Gene presence does not guarantee phenotypic resistance
  • Novel mechanisms: Databases lag behind emerging resistance determinants

The Series describes fragmented data and algorithmic-bias barriers in high-income settings, alongside limited infrastructure, nonstandardized data, and financial constraints in many low-income and middle-income settings. A model developed in one resistance ecology should not be transferred to another without local and temporal validation.

Structured Data vs. Clinical Notes for AMR Prediction

Recent evidence challenges assumptions that clinical narratives will always improve AMR prediction. In a 2026 conference abstract describing a retrospective cohort across 10 hospitals, the structured-data model achieved AUROC 0.85 for community-onset resistant Gram-negative sepsis, compared with 0.74 for the best note-based LLM; combining approaches did not significantly improve prediction (Hixon et al., 2026, conference abstract). The abstract does not establish prospective performance or clinical benefit.

In this dataset and modeling comparison, structured clinical history carried more useful predictive signal than the evaluated note-based LLMs. That result should not be generalized to every pathogen, note type, task, or model family.

Part 3: Healthcare-Associated Infection Surveillance

AI in Infection Prevention

Infection preventionists face increasing surveillance requirements with limited staffing. AI promises to automate case finding and risk stratification.

Digital Surveillance Evidence Base

Digital Infection-Prevention Evidence and Requirements

SHEA-affiliated authors have described the data requirements and validation problems of electronic HAI surveillance (Woeltje et al., 2014). The following checklist is an implementation synthesis, not a formal society endorsement of specific AI applications:

Candidate applications:

  • Automated HAI surveillance from EHR data
  • Real-time syndromic surveillance dashboards
  • Hand hygiene monitoring systems (with privacy considerations)
  • Antimicrobial use tracking and benchmarking

Implementation requirements:

  • Validation against manual chart review (sensitivity, specificity, PPV)
  • Transparent algorithms with audit capability
  • Integration with existing infection control workflows
  • Privacy protections for patient and staff data

Cautions:

  • AI predictions should not replace clinical judgment for isolation decisions
  • False positives consume investigation resources
  • Equity review required (do algorithms perform equally across patient populations?)

HAI Risk Prediction Evidence

Infection Type Model Performance Clinical Impact
Surgical site infection Risk prediction or automated case finding External validation and action-pathway evidence
CAUTI Criteria-based surveillance or risk prediction Separation of colonization, device exposure, and infection
CLABSI Case finding, device-day tracking, or risk prediction Validation against current definitions and manual review
C. difficile infection Testing, isolation, or recurrence support Reference-label quality and consequences of false positives

Performance across settings: The 2026 Lancet Infectious Diseases review describes infrastructure, data-standardization, interoperability, and bias barriers that differ across health systems (Miglietta et al., 2026). It does not establish one universal percentage loss in sensitivity after transfer. Models require local validation before deployment, particularly where pathogen distributions, definitions, documentation, and infection-control resources differ.

Key challenge: Predicting HAI is inherently difficult because it requires distinguishing colonization from infection and accounting for interventions that occur after prediction.

Automated Surveillance Systems

Commercial systems (Premier, Theradoc, Vigilanz) incorporate:

  • NHSN criteria-based case finding
  • Antibiotic exposure tracking
  • Line and device day counting
  • Automated reporting generation

Evidence boundary: The Woeltje white paper reviews data requirements and validation considerations for electronic HAI surveillance; it does not support one universal 85%–95% sensitivity range for commercial systems. Each implementation should report case-definition version, data completeness, sensitivity, specificity, positive predictive value, adjudication workload, and discrepancies against trained manual review.

Part 4: Rapid Diagnostics and Resistance Prediction

Laboratory Automation and AI-Enhanced Microbiology

MALDI-TOF Mass Spectrometry:

Matrix-assisted laser desorption ionization-time of flight (MALDI-TOF) mass spectrometry changed clinical microbiology by matching spectral profiles to reference databases. Established species-identification workflows are automated pattern matching and should not automatically be labeled machine learning (Seng et al., 2009).

Current performance is organism-, specimen-, preparation-, database-, threshold-, and platform-specific. Commercial systems include VITEK MS and MALDI Biotyper, but product configuration and regulatory status vary by version and market. Unusual species, mixed cultures, and incomplete reference libraries remain important limitations.

Gram Stain AI:

Computer vision analysis of Gram stain images:

  • Report the exact morphology task, organism mix, slide preparation, reference labels, and external test set rather than a transferred accuracy range
  • Challenges: slide quality variation, mixed flora, unusual organisms
  • Limited clinical deployment; most systems in validation phase

Resistance Prediction from Spectral Data:

Emerging research uses MALDI-TOF spectra to predict: - MRSA vs. MSSA - Carbapenemase production - Extended-spectrum beta-lactamase (ESBL)

Transportability, not internal discrimination, is the central limitation. In a corrected validation study, models trained on German or Swiss DRIAMS spectra performed poorly when location and time differed, and performance declined during 18 months of prospective observation. The authors concluded that local training and regular retraining were necessary (Wiesmann et al., 2025; Wiesmann et al., correction, 2026). In the publisher’s unedited early-access version of the ANTIBIOTIC study, median AMR AUC fell from 0.94 internally to 0.55 on a temporal test set and improved only to 0.61 after fine-tuning with recent data (Wang et al., 2026). The paper reports integration of a large language model recommendation layer but reports no separate clinical validation of that layer. These models do not replace reference AST.

A 2026 systematic review of 57 K. pneumoniae AI studies found high overall risk of bias in 45 studies, external validation in 17, and prospective real-world evaluation in only one. MALDI-TOF was the most common input modality (Aggarwal et al., 2026). High internal AUCs in retrospective spectral datasets do not establish clinical readiness.

Laboratory Workflow Tools

MALCA: MALCA uses disk-diffusion inhibition zones to classify carbapenemase-producing Enterobacterales and predict enzyme class. A blinded, temporally independent cohort was evaluated, but all antibiograms used one center’s workflow and one reagent setup. The authors call for prospective multi-site validation, and the public repository contains an illustrative workflow rather than the restricted full codebase and trained models (Emeraud et al., 2026). The publication therefore does not provide a deployable clinical package.

Antibiogo: Antibiogo is an offline Android tool that guides measurement and interpretation of disk-diffusion AST. Its peer-reviewed technical validation assessed agreement with automated and manual reading, not patient outcomes (Pascucci et al., 2021). MSF reports routine deployment in its laboratories and IVDD-CE marking for the EUCAST configuration (MSF Foundation, accessed July 2026; Antibiogo, September 2025). The current official product site routes access through institutional early-access registration (Antibiogo, accessed July 2026). Antibiogo supports reading an already-performed susceptibility test; laboratory quality control and clinical interpretation remain necessary.

Molecular Diagnostics and AI

Syndromic panels such as BioFire and Verigene use automated assay interpretation. They are important diagnostic inputs to stewardship, but multiplex detection and rule-based reporting are not necessarily AI:

  • Multiplexed pathogen detection
  • Automated resistance gene reporting
  • Integration with stewardship pathways

Clinical impact: The Ramanan review describes faster organism and resistance-marker information from multiplex molecular panels and emphasizes that impact depends on diagnostic stewardship and how clinicians act on results (Ramanan et al., 2018). It does not support one universal 24–48-hour improvement across panels and workflows.

Imaging-based WHO–IWGE classification of hepatic cystic echinococcosis

WHO–IWGE CE1–CE5 subtype determines whether hepatic cystic echinococcosis is managed with albendazole, PAIR, or surgery, but expert staging is uneven in endemic hospitals that often have only non-contrast CT. A 17 August 2026 npj Digital Medicine multicentre retrospective study trained a two-stage multimodal model (lesion segmentation then CE1–CE5 classification, fusing 2D, 3D, and radiomics features) on 733 patients from Qinghai University Affiliated Hospital and externally validated it in 261 patients from two county-level hospitals (Wang, Z. et al., 2026). The multimodal model’s macro-AUC was 0.942 (95% CI 0.910–0.964) internally and 0.931 (0.898–0.956) externally; most errors were between adjacent transitional stages. In the paper’s reader study, AI-assisted interpretation raised observed macro-AUC versus physician-only reads in both the tertiary cohort (0.970 vs 0.905) and the county-hospital cohort (0.963 vs 0.902). This is a retrospective discrimination and reader-study result on NCCT, not a prospective treatment-allocation or outcome trial, and it does not replace expert WHO–IWGE staging or justify autonomous CE-class assignment.

Part 5: Outbreak Prediction and Response

What Works: Nowcasting

Nowcasting estimates current disease burden from incomplete, lagged data. Forecasting systems include statistical, mechanistic, ensemble, and machine-learning approaches:

  • CDC FluSight: Ensemble models evaluate multiple short-horizon influenza targets
  • COVID-19 Forecast Hub: Aggregated models for hospitalizations and deaths
  • Performance should be reported for the exact target, place, horizon, scoring rule, and evaluation period; uncertainty often grows with horizon

Key insight: Nowcasting can reduce the operational effect of reporting delay, but its value must be demonstrated against the surveillance data and decisions already available.

What Does Not Work: Long-Range Prediction

Current evidence does not support reliable prediction of:

  • When and where novel outbreaks will emerge
  • Pandemic timing or severity
  • Long-range seasonal disease patterns

Why prediction fails:

  • Rare events (low base rate problem)
  • Complex, stochastic determinants (human behavior, weather, pathogen evolution)
  • Data quality and timeliness limitations
  • Non-stationary dynamics (patterns change over time)

The Conceptual Framework for Prediction Failure

The 2026 Lancet Infectious Diseases conceptual framework maps AI applications across pathogens, hosts, environments, research, public health, and clinical practice, while emphasizing the available evidence and technical, ethical, policy, and implementation barriers (Odone et al., 2026). It does not establish that one model class can solve outbreak emergence prediction.

Recurrent sources of forecast uncertainty:

  1. Sparse emergence events: Rare outcomes provide limited examples for model development and prospective evaluation
  2. Behavioral feedback: Interventions and public responses change the trajectory being forecast
  3. Pathogen evolution: New variants and recombination can alter transmissibility, immune escape, or severity
  4. Data latency and revision: Case, hospitalization, laboratory, and genomic data arrive with different delays and are revised over time

The defensible claim is target-specific: a model may improve a defined nowcast or short-horizon forecast without being able to predict novel emergence. Long-horizon claims require prospective, repeated evaluation and calibrated uncertainty rather than confidence based on one outbreak.

The practical implication for ID specialists: invest in surveillance infrastructure that enables rapid response rather than prediction systems that promise early warning.

Syndromic Surveillance

Automated and machine-learning-supported syndromic surveillance can include:

  • Emergency department chief complaint monitoring
  • OTC medication sales tracking
  • Social media and search trend analysis

Performance: Detection depends on baseline incidence, reporting coverage, signal-to-noise ratio, threshold, geography, and outbreak size. Systems should report event-level sensitivity, false-alert burden, detection timing, and missed outbreaks rather than a general claim that anomaly detection works.

Part 6: Professional Society Positions

IDSA Position on Digital Health

IDSA on Digital Health and Telemedicine

The Infectious Diseases Society of America has addressed digital tools through position statements on telemedicine (Young et al., 2019):

Telemedicine and digital consultations:

  • Remote ID consultation supports stewardship in facilities without on-site ID
  • Telehealth platforms can provide educational opportunities, share tools, and allow antimicrobial use review with feedback
  • Documentation and liability considerations apply to technology-assisted consultations

Antimicrobial stewardship implications:

  • AI decision support should align with IDSA/SHEA stewardship guidelines
  • Local validation required before deployment
  • Human oversight of AI recommendations is essential

Genomic diagnostics:

  • WGS-based diagnostics require interpretive expertise
  • Automated interpretation tools should not replace ID physician review
  • Turnaround time for genomic results must meet clinical needs

Note: IDSA has not released a formal position statement specifically on AI. The principles above reflect general guidance applicable to clinical decision support.

CDC and WHO Positions

The CDC Core Elements of Hospital Antibiotic Stewardship Programs recommends accountable stewardship infrastructure and supports decision-support approaches, but it is not an AI validation standard.

The World Health Assembly adopted the updated WHO Global Action Plan on Antimicrobial Resistance, 2026–2036, a One Health policy framework rather than an AI validation standard. The WHO AWaRe antibiotic book provides evidence-based antibiotic guidance that may be delivered through CDS; the recommendation and any predictive model layered onto it require separate evaluation.

Hypothetical Clinical Scenarios

The following fictional cases are decision exercises. They do not report actual patients, product performance, or validated local thresholds.

Hypothetical case: A 68-year-old man with bacteremia is receiving piperacillin-tazobactam. Blood cultures grow E. coli reported susceptible to ceftriaxone, ampicillin-sulbactam, and trimethoprim-sulfamethoxazole. A stewardship system generates a de-escalation alert for the attending physician.

The alert states: “Current therapy: piperacillin-tazobactam. Organism susceptible to narrower-spectrum agents. Consider de-escalation to ceftriaxone.”

Question: How should the ID consultant approach this alert?

Discussion

Key considerations:

  1. Source control: Has the source of bacteremia been identified and addressed?
  2. Clinical trajectory: Is the patient improving on current therapy?
  3. Polymicrobial risk: Are cultures from other sites pending?
  4. Patient factors: Allergies, prior resistant organisms, immunocompromise?

The AI’s limitation: The alert is based solely on the susceptibility report. It lacks context about clinical status, imaging findings, or pending workup.

Appropriate response:

  • Review the chart before acting on the alert
  • Document rationale if continuing broad-spectrum therapy
  • De-escalate when clinically appropriate, not simply because of an alert
Teaching point: Stewardship AI identifies opportunities for de-escalation. The decision to de-escalate requires clinical judgment that integrates information the AI does not have.

Hypothetical case: A 45-year-old woman with recurrent urinary tract infections has a positive urine culture. The laboratory performs rapid molecular testing that detects blaCTX-M. Phenotypic susceptibilities are pending.

The EHR displays: “ESBL gene detected. Consider carbapenem therapy pending phenotypic susceptibilities.”

Question: How should this information guide empiric therapy?

Discussion

Genotype-phenotype considerations:

  1. Gene detection supports an ESBL mechanism but assay scope, expression, organism, and other mechanisms still matter
  2. Carbapenems are reliable for ESBL-producers but represent overtreatment if alternatives work
  3. Cephalosporins may fail but MIC testing will clarify
  4. Nitrofurantoin and fosfomycin may have activity regardless of ESBL status

The nuanced approach:

  • For severe infection: carbapenem empirically while awaiting phenotypic susceptibilities
  • For uncomplicated cystitis: can await full susceptibilities given low severity
  • Review prior cultures and susceptibility patterns
Teaching point: Genotypic resistance detection is faster than phenotypic testing but requires interpretation. The presence of a resistance gene informs but does not dictate therapy.

Hypothetical case: The infection-prevention team identifies three patients with carbapenem-resistant K. pneumoniae on the same unit over two weeks. The hospital’s validated WGS pipeline reports that the isolates differ by five single-nucleotide polymorphisms.

Question: What does the genomic data contribute to the investigation?

Discussion

Interpretation of genetic distance:

  • A five-SNP difference may support close relatedness within a defined pipeline and background population
  • No universal SNP cutoff proves direct transmission or direction
  • The finding should be integrated with sampling density, dates, locations, epidemiologic links, and within-host diversity

What genomics does not tell you:

  • Who was the index case
  • The specific transmission route (hands, environment, devices)
  • Whether there are undetected carriers

Investigation steps:

  1. Enhanced contact precautions for known cases
  2. Review common exposures (procedures, staff, equipment)
  3. Consider point prevalence surveillance for additional carriers
  4. Environmental cultures if transmission route unclear
Teaching point: Genomic epidemiology can strengthen or weaken a transmission hypothesis. It does not independently prove direct transmission, direction, or route, and it does not replace field epidemiology.

Hypothetical case: An ID consultant receives a page because a sepsis model alert fired on a 72-year-old patient with pneumonia on the medicine floor. The patient’s vital signs are stable, and lactate is normal. The alert cites an elevated white count and recent antibiotic initiation as risk factors.

Question: How should the ID consultant interpret and respond to this alert?

Discussion

Context on sepsis prediction AI:

Sepsis prediction models have significant limitations (see Critical Care chapter): - False-positive burden and positive predictive value vary by model, threshold, population, and response workflow - The Epic Sepsis Model, widely deployed, showed 33% sensitivity and 83% specificity in external validation (Wong et al., 2021)

This specific case:

  • Stable vital signs and normal lactate are reassuring
  • Antibiotic initiation and elevated WBC are expected for pneumonia treatment
  • The alert adds little to clinical assessment

Appropriate response:

  1. Brief chart review to confirm clinical stability
  2. No action required based solely on the AI alert
  3. Continue planned pneumonia management
  4. Document assessment if institution requires alert acknowledgment
Teaching point: A sepsis alert should trigger a proportionate clinical assessment defined by the local protocol. The model output neither proves sepsis nor justifies ignoring objective deterioration.

Key Takeaways for ID Specialists

Practical Guidance

Antimicrobial stewardship AI:

  • AI tools identify de-escalation opportunities; clinical judgment determines appropriateness
  • Prospective implementations can improve prescribing process measures without establishing patient-outcome benefit
  • Local validation essential: national resistance data may not reflect your institution
  • Alert fatigue is real: advocate for tiered, high-value alerts

Genomic surveillance:

  • Phylogenetic analysis confirms or refutes transmission clusters
  • Automated lineage assignment enables rapid variant tracking
  • Genotypic resistance prediction is fast but requires phenotypic confirmation

Outbreak prediction:

  • Nowcasting and short-horizon forecasts can support defined operational decisions when prospectively evaluated
  • Performance and uncertainty usually change with horizon, target, location, and reporting conditions
  • AI augments surveillance; it does not predict novel emergence

HAI prevention:

  • Automated surveillance improves case finding sensitivity
  • Risk prediction models require prospective validation
  • Alert fatigue affects infection prevention as it does other clinical domains

Professional standards:

  • IDSA/SHEA guidelines apply to AI-assisted stewardship
  • Human oversight of AI recommendations is required
  • Reference AST and confirmatory pathways remain necessary for resistance-prediction systems
  • Monitor resistance models for site and temporal drift before considering retraining under controlled validation
  • Local validation and equity review are essential before deployment

Questions About AI in Infectious Diseases

How effective is AI for antimicrobial stewardship?

AI stewardship evidence is heterogeneous and is dominated by retrospective prediction and observational implementation studies. Benefit depends on the model, local resistance ecology, stewardship workflow, comparator, and measured endpoint; general drug-alert override rates should not be presented as AI-specific efficacy evidence.

Can AI predict disease outbreaks?

Models can estimate current burden and produce short-horizon forecasts, but performance varies by pathogen, location, target, horizon, reporting delay, and behavioral change. No model can reliably predict the timing and location of novel outbreak emergence.

What is Nextstrain and how does it use AI?

Nextstrain is an open-source genomic-epidemiology platform for reproducible phylogenetic analysis and visualization. Its pipelines are automated computational tools, but use of Nextstrain is not by itself evidence that a machine-learning model improved surveillance or patient outcomes.

Can AI predict antimicrobial resistance from laboratory or genomic data?

Genotype-to-phenotype and spectral resistance prediction are promising, but transportability varies across sites and time. These systems require local and temporal validation and do not replace reference antimicrobial susceptibility testing.