The review, led by Yi Peng and colleagues at The First People’s Hospital of Guangyuan in Sichuan, China, examines how biomarkers, continuous physiological monitoring, electronic patient-reported outcomes, and caregiver-reported behavioral changes can be fused by AI into individualized risk trajectories for children undergoing immunotherapy. The authors propose a three-layer framework linking immune-toxicity biology, a multimodal digital phenotype, and what they call nursing counter-regulation — the structured clinical response that converts algorithmic signals into timely bedside action.

What makes pediatric immunotherapy monitoring so different from conventional chemotherapy surveillance is the speed, nonlinearity, and overlap of immune-mediated toxicities. CAR-T cells and T-cell engagers can trigger cytokine release syndrome (CRS), a systemic inflammatory cascade involving immune-cell expansion, cytokine release, endothelial activation, capillary leak, and organ dysfunction. CRS grading is driven by fever, hypotension, and hypoxia. A related toxicity, immune effector cell-associated neurotoxicity syndrome (ICANS), requires structured neurologic assessment and may include language disturbance, impaired attention, altered consciousness, motor findings, seizures, or cerebral edema. In younger children or those with pre-existing neurodevelopmental differences, ICANS can be difficult to recognize before overt deterioration.

Biomarkers such as C-reactive protein, ferritin, interleukin-6, interferon-gamma, angiopoietin-2, and von Willebrand factor have been associated with severe CRS or neurotoxicity, but none is sufficiently specific to function independently of clinical assessment. A later hyperinflammatory phenotype, immune effector cell-associated hemophagocytic lymphohistiocytosis-like syndrome, adds rapidly rising ferritin, cytopenias, coagulopathy, and hepatic injury. Meanwhile, checkpoint inhibitors produce a different surveillance problem altogether: immune-related adverse events may affect the skin, gastrointestinal tract, liver, endocrine organs, lungs, heart, kidneys, nervous system, or hematopoietic system, and may arise during treatment or months after discontinuation. Anti-GD2 antibodies can add severe neuropathic pain, hypersensitivity, capillary leak, and autonomic changes.

The same observed signal can have competing explanations across modalities. Tachycardia may reflect pain, dehydration, sepsis, or CRS. New irritability may reflect anxiety, corticosteroids, sleep deprivation, or emerging neurotoxicity. Somnolence may reflect medication, sleep disruption, metabolic disturbance, or ICANS. Fever may represent infection, CRS, or both. Monitoring must therefore be longitudinal, multimodal, and context aware rather than dependent on a single threshold — precisely the kind of problem where AI’s capacity to integrate fragmented observations into a time-dependent estimate of risk becomes mechanistically valuable.

The review’s authors are explicit that the clinically relevant mechanism is not autonomous prediction. It is a nurse-in-the-loop system in which AI improves temporal sensitivity and prioritization, while nurses verify signal quality, interpret age- and development-specific context, reconcile child–caregiver discordance, activate protocol-based escalation, and maintain accountability. A multimodal model could combine treatment timing, patient-specific baseline values, laboratory trajectories, bedside and wearable vital signs, medication exposures, nursing observations, electronic patient-reported outcomes, and caregiver-reported behavioral changes. Time-series methods and anomaly detection are particularly relevant because an individual child’s deviation from baseline may be more informative than a population threshold. For example, a small but sustained rise in respiratory rate coupled with declining oxygen saturation, increasing fatigue, and a rising inflammatory marker may warrant review even when no single value is alarming.

Before treatment, nurses should establish the baseline against which future AI outputs will be interpreted. This includes usual heart rate and blood pressure ranges, oxygen requirement, sleep and activity pattern, hydration status, bowel habits, skin findings, pain phenotype, gait, speech, attention, handwriting or age-appropriate cognitive tasks, and the child’s typical response to hospitalization. Baseline laboratory values should be linked to recent transfusion, infection, growth-factor exposure, corticosteroids, organ dysfunction, and prior therapy. A model that ignores these modifiers may generate false escalation or falsely reassure. The nursing assessment should also define the family’s monitoring capacity: caregivers need a clear description of expected symptoms, red flags, who will review electronic reports, how rapidly the team will respond, and when emergency evaluation is required regardless of the AI-generated risk score.

During the high-risk period after CAR-T or T-cell engager administration, nurses remain the first continuous interpreters of clinical change. Core observations include temperature, heart rate, blood pressure, respiratory rate, oxygen saturation, perfusion, urine output, fluid balance, weight, pain, mobility, mental status, speech, attention, tremor, seizure activity, and medication exposure. An AI alert should prompt a structured reassessment rather than automatic treatment. AI-generated risk scores should determine attention priority, not toxicity grade; formal CRS and ICANS grading remains based on ASTCT criteria and bedside findings. Verification begins with signal quality: nurses determine whether an abnormal wearable value is physiologic or caused by poor contact, movement, anxiety, or device failure. They repeat vital signs, review the timing of antipyretics, sedatives, opioids, transfusions, corticosteroids, and antihypertensives, and examine whether multiple domains are changing together. Infection and CRS may coexist; cultures and antimicrobial pathways should not be delayed because an inflammatory model labels the event as immunotoxicity.

Alert fatigue is a direct patient-safety risk. Systems should suppress duplicate notifications, rank alerts by urgency, display the variables driving the score, and show uncertainty. A single unexplained risk number is inadequate. Nurses need to know whether the alert reflects worsening oxygenation, a ferritin trajectory, new caregiver-reported confusion, or missing data. Escalation rules should be tested during simulation, including night shifts and low-staffing periods, before clinical deployment. When bedside nursing concern is high despite a low predicted risk, clinical assessment should override the algorithm and trigger reassessment or escalation.

Subjective toxicities are frequently underrecognized when assessment relies only on clinician documentation. The Pediatric Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events has demonstrated validity and reliability in children and adolescents, and symptom screening tools such as SSPedi can improve symptom detection and trigger care pathways. Electronic monitoring is feasible during pediatric cancer hospitalization and in weekly patient- or proxy-reported programs. Age and development determine who reports and how the report should be interpreted: older children may self-report pain, fatigue, dyspnea, cognitive slowing, and functional interference, whereas younger children may require caregiver proxy reporting and behavioral anchors. Discordance between the child and caregiver is not an error to be averaged away — a child reporting mild symptoms while a caregiver reports marked behavioral change may indicate adaptation, fear of hospitalization, or emerging neurotoxicity. Nurses are positioned to reconcile these perspectives through focused questioning and direct observation. Within the proposed AI framework, child and proxy reports should remain distinguishable inputs; the magnitude and direction of discordance can itself be treated as a clinically informative feature or trigger for nursing reassessment.

After discharge, the surveillance model changes from continuous bedside observation to intermittent family-generated data. Wearable devices can record multiparameter vital signs in pediatric oncology, demonstrating the technical feasibility of high-frequency physiological monitoring. Daily electronic symptom reporting has also been implemented in children with cancer. For immunotherapy, the clinically useful home system should combine simple measurements with a nurse-led triage pathway: temperature, oxygen saturation when indicated, activity and sleep trends, oral intake, urine output, pain, diarrhea, rash, cough, neurologic or behavioral change, and medication adherence. Thresholds must reflect time from therapy, prior toxicity, distance from care, and the reliability of the home data source. A low AI-generated risk score must never prevent families from seeking urgent care. Neurologic change, respiratory compromise, hemodynamic instability, or persistent fever requires urgent clinical assessment independent of wearable or electronic output.

Long-term monitoring must extend beyond the period conventionally associated with CRS or ICANS. Nurses should coordinate surveillance for B-cell aplasia, immunoglobulin deficiency, recurrent infection, delayed vaccination, prolonged cytopenia, fatigue, neurocognitive change, school reintegration, endocrine dysfunction, cardiopulmonary symptoms, growth, puberty, fertility concerns, emotional distress, caregiver burden, and financial toxicity. Checkpoint inhibitor immune-related adverse events may persist or first appear after therapy has ended, while the cumulative effects of prior chemotherapy, radiation, transplantation, and immunotherapy can be difficult to disentangle. AI may support survivorship by assembling longitudinal laboratory and symptom trajectories, identifying missed monitoring, and generating risk-adjusted reminders. It could also flag a pattern that crosses specialties — for example, declining activity, school absence, persistent tachycardia, and endocrine symptoms. The nurse remains responsible for determining whether the signal is clinically meaningful, arranging targeted assessment, coordinating referrals, and explaining the plan to the child and family.

AI-enabled monitoring will fail if alerts are detached from nursing communication. Each handover should preserve the treatment day, current risk window, highest prior CRS or ICANS grade, recent biomarker direction, organ-specific symptoms, antimicrobial and immunosuppressive exposure, and the child’s developmental baseline. The receiving nurse should be able to see which alerts occurred, what bedside verification found, what action was taken, and when reassessment is due. A closed-loop status — acknowledged, assessed, escalated, resolved, or deferred with rationale — prevents alerts from disappearing into an electronic inbox. Nursing quality review should examine more than model sensitivity: useful indicators include time from alert to bedside reassessment, time to clinician escalation, duplicate-alert rate, overnight response, false reassurance, family-initiated urgent contact, unplanned intensive care, and the proportion of alerts without documented closure.

The evidence base for AI specifically applied to pediatric immunotherapy nursing remains early. Most available studies address general clinical prediction, adult oncology, or feasibility of pediatric digital monitoring rather than prospective improvement in immunotherapy outcomes. Broad claims of automated prediction should therefore be avoided. Machine learning can identify complex patterns, but performance depends on data quality, label definition, sample size, and transportability. Pediatric datasets are small, treatment protocols evolve, and severe toxicities are relatively infrequent; models may overfit local practice or fail when transferred to another center. Clinically evaluated pediatric digital approaches currently include electronic symptom monitoring and wearable feasibility studies, whereas multimodal AI fusion and prospective immunotherapy-specific toxicity forecasting remain largely proposed.

Bias can arise from unequal device access, missing electronic reports, language differences, age-dependent communication, skin-tone effects on sensors, and institutional variation in laboratory timing. A model can appear accurate while systematically underperforming for children who generate less digital data. Algorithmic bias in health care is not theoretical, and subgroup calibration must be reported. Privacy and cybersecurity are especially important because pediatric data may include continuous location-linked physiology, family narratives, and long-term developmental information. Consent and assent processes should explain what is collected, who reviews alerts, how long data are retained, and whether data will be used to update models.

A practical research agenda should begin with silent prospective validation, followed by nurse-supervised clinical pilots and multicenter pragmatic evaluation. Models should be recalibrated as products, protocols, and supportive care change. Reporting should follow AI-specific guidance, and clinical governance should require an auditable override mechanism, downtime procedures, and accountability for missed or excessive alerts. Nurses should participate in model design, label adjudication, usability testing, and governance because they understand which data are reliable, when documentation lags behind reality, and how alerts alter workload. Outcome measures should include time to bedside reassessment, time to escalation, severe toxicity, unplanned intensive care, symptom burden, family-reported confidence, nursing workload, alert burden, and equity — not only area under the receiver operating characteristic curve.

AI-enabled surveillance can strengthen pediatric cancer immunotherapy care by linking biological markers, continuous physiology, nursing observations, and the child’s and caregiver’s reported experience. Its most plausible clinical mechanism is improved temporal sensitivity: a model identifies a changing multimodal risk state, and a nurse verifies the signal, supplies developmental and treatment context, activates standardized reassessment, and coordinates intervention. The technology should therefore be evaluated as a nurse-in-the-loop safety system, not as an autonomous diagnostic tool. Future studies must demonstrate that AI improves these nursing processes and patient-centered outcomes without increasing alert fatigue, inequity, or documentation burden. Until such evidence is available, the safest principle is augmentation with accountability: AI may prioritize attention, but pediatric oncology nurses remain the clinical integrators who convert data into timely, family-centered care.

Subject of Research: Artificial intelligence–enabled nursing surveillance for pediatric cancer immunotherapy: integrating biomarkers, patient-reported outcomes, and long-term toxicity management

Article Title: Artificial intelligence–enabled nursing surveillance for pediatric cancer immunotherapy: integrating biomarkers, patient-reported outcomes, and long-term toxicity management

Article References: Peng, Y., Kou, Q., Guo, S., Xu, J., Wang, L., & Chen, X. (2026). Artificial intelligence–enabled nursing surveillance for pediatric cancer immunotherapy: integrating biomarkers, patient-reported outcomes, and long-term toxicity management. Frontiers in Medicine, 13, Article 1976674. https://doi.org/10.3389/fmed.2026.1976674

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1976674

Keywords: Artificial, intelligence, enabled, nursing, surveillance, pediatric, cancer, immunotherapy, integrating, biomarkers, patient-reported, outcomes