From virtual reality to artificial intelligence-driven digital humans in perioperative lung cancer care: An updated scoping review and translational framework for pulmonary rehabilitation

Abstract

Digital pulmonary rehabilitation may improve access and continuity around lung cancer surgery, but virtual reality, scripted avatars, connected applications, and artificial intelligence (ai)-driven digital humans are frequently conflated. we conducted an updated scoping synthesis using a two-stage architecture: an inherited systematic review corpus searched through January 1, 2024, and a targeted PubMed/citation update from January 2, 2024, to August 31, 2026. Direct clinical studies were classified by operational Level 1–3 criteria. Risk of bias (RoB) was appraised with RoB 2 or ROBINS-I where feasible; implementation findings were mapped to Nonadoption, Abandonment, Scale-up, Spread, and Sustainability while separating reported barriers from framework-derived concerns. The direct map contained 19 completed reports and one protocol: 5 Level 1 fixed-display/immersive interventions, 14 Level 2 deterministic connected systems, no completed Level 3 digital human trials, and one Level 2/3 protocol candidate. In the inherited meta-analysis, digital rehabilitation was associated with fewer postoperative pulmonary complications (risk ratio: 0.45, 95% confidence interval [CI]: 0.30–0.66) and shorter chest tube duration (−1.53 days), whereas postoperative length of stay was not significantly reduced (−1.42 days, 95% CI: −3.45 to 0.62). Updated studies reported heterogeneous functional and workflow benefits, but no direct evidence validated Level 3 outcome superiority. Current evidence supports digitally delivered rehabilitation, not AI-driven digital humans. The proposed taxonomy converts this evidence gap into testable design requirements. Future trials should prioritize head-to-head comparisons, standardized outcomes, AI safety and external validation, geriatric accessibility, interoperability, health economics, and sustained real-world implementation.

Background

Lung cancer remains the leading cause of cancer mortality worldwide.[1] Surgical resection is central to curative treatment for early-stage disease, but postoperative pulmonary complications (PPCs), functional decline, and fragmented postdischarge support continue to undermine recovery. Enhanced recovery after surgery (ERAS) recommendations therefore emphasize education, pulmonary care, and early mobilization.[2,3]

Digital pulmonary rehabilitation can extend exercise instruction, symptom monitoring, and feedback beyond the hospital. A 2024 systematic review identified 11 perioperative digital rehabilitation studies and suggested benefits for selected outcomes, but also documented marked heterogeneity, incomplete intervention reporting, and high risk of bias (RoB).[4] A broader review of supportive eHealth in lung cancer likewise found heterogeneous technologies and outcomes.[16]

Terminology has advanced faster than the evidence. Fixed videos, virtual reality (VR), rule-based applications, chatbots, and generative systems are often described as avatars, coaches, or digital humans despite materially different adaptability, sensing, and clinical risk. This ontological ambiguity can cause evidence from conventional digital delivery to be incorrectly attributed to artificial intelligence (AI). We therefore conducted an updated scoping review and translational framework to: (1) update the direct perioperative lung-cancer evidence; (2) apply reproducible operational rules for three technology levels; (3) separate evidence density, methodological quality, and certainty; and (4) define a prioritized pathway for safe evaluation of AI-driven digital humans.

Materials and Methods

Review design and evidence architecture

This work was structured as an updated scoping review and translational framework in accordance with joanna briggs institute (JBI) methodological guidance and Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR).[5,6] It was not designed as a new meta-analysis. Where pooled estimates are reported, they are reproduced from the source systematic review rather than recalculated.[4] The completed PRISMA-ScR reporting checklist is provided in Supplementary Table 1.

Supplementary Table 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews reporting checklist (manuscript mapping).

PRISMA-ScR item		How addressed		Location
Title		Identifies an updated scoping review and translational framework		Title
Structured summary		Background, methods, results, conclusion		Abstract
Rationale/objectives		Ontological gap and four objectives		Background
Protocol/registration		No prospectively registered protocol documented; limitation stated		Methods; Study limitations
Eligibility		PCC criteria and exclusions		Eligibility criteria
Information sources		Inherited corpus plus targeted PubMed/citation update		Review design
Search		Exact PubMed and reproducible database strategies		Supplementary Table 1
Selection/data charting		Single-reviewer targeted update and verification stated		Eligibility/study selection
Critical appraisal		Inherited and provisional updated appraisals		Methods; Supplementary Table 4
Synthesis		Descriptive mapping by technology level/outcome		Results; Tables 1–3; Figures 1–2
Study characteristics		Direct evidence map		Supplementary Table 3
Limitations		Search, screening, heterogeneity, inference, Level 3 gap		Study limitations
Conclusions		Evidence limited to Levels 1–2		Conclusion

PRISMA-ScR=Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews

The evidence architecture had two stages. Stage 1 retained the 11-study corpus of the source review, which searched six English-language and four Chinese-language databases from inception to January 1, 2024.[4] Stage 2 was a targeted update of PubMed/MEDLINE plus backward and forward citation searching from January 2, 2024, to August 31, 2026. Search concepts combined lung cancer, surgery, pulmonary rehabilitation or exercise, and digital modalities; the exact PubMed strategy and reproducible strategies for Embase, Web of Science, Scopus, CINAHL, and CENTRAL are provided in Supplementary Table 2. The update was limited to English-language indexed reports. Because the submitted project archive did not contain a de novo multidatabase export or duplicate-screening log, the update is explicitly reported as a targeted, single-reviewer verification rather than a fully duplicated systematic rerun.

Supplementary Table 2.

Search architecture and database-specific strategies.

Database		Strategy
PubMed/MEDLINE		((“lung neoplasms”[Mesh] OR “lung cancer”[tiab] OR “lung cancers”[tiab] OR “pulmonary neoplasm”[tiab] OR “pulmonary neoplasms”[tiab] OR NSCLC[tiab]) AND (surg*[tiab] OR resection[tiab] OR lobectomy[tiab] OR segmentectomy[tiab] OR thoracoscop*[tiab] OR perioperat*[tiab]) AND (“pulmonary rehabilitation”[tiab] OR prehabilitation[tiab] OR telerehabilitation[tiab] OR “respiratory training”[tiab] OR “breathing exercise”[tiab] OR “breathing exercises”[tiab] OR exercise[tiab]) AND (“digital health”[tiab] OR telehealth[tiab] OR mHealth[tiab] OR “mobile app”[tiab] OR “mobile apps”[tiab] OR smartphone*[tiab] OR wearable*[tiab] OR exergam*[tiab] OR “virtual reality”[tiab] OR avatar*[tiab] OR chatbot*[tiab] OR “conversational agent”[tiab] OR “conversational agents”[tiab] OR “digital human”[tiab] OR “digital humans”[tiab] OR “digital therapeutic”[tiab] OR “digital therapeutics”[tiab] OR “artificial intelligence”[tiab] OR “internet of things”[tiab])) AND (“2024/01/02”[Date - Publication] : “2026/08/31”[Date - Publication])
Embase		(“lung cancer”/exp OR “lung cancer”:ti,ab OR nsclc:ti,ab) AND (“lung surgery”/exp OR surg*:ti,ab OR resection:ti,ab OR lobectomy:ti,ab OR segmentectomy:ti,ab OR thoracoscop*:ti,ab OR perioperat*:ti,ab) AND (“pulmonary rehabilitation”/exp OR prehabilitation:ti,ab OR telerehabilitation:ti,ab OR “respiratory training”:ti,ab OR “breathing exercise*”:ti,ab OR exercise:ti,ab) AND (“digital health”/exp OR telehealth:ti,ab OR mhealth:ti,ab OR “mobile app*”:ti,ab OR smartphone*:ti,ab OR wearable*:ti,ab OR exergam*:ti,ab OR “virtual reality”:ti,ab OR avatar*:ti,ab OR chatbot*:ti,ab OR “conversational agent*”:ti,ab OR “digital human*”:ti,ab OR “digital therapeutic*”:ti,ab OR “artificial intelligence”:ti,ab OR “internet of things”:ti,ab) AND [2024-2026]/py
Web of Science Core Collection		TS=((“lung cancer” OR “pulmonary neoplasm*” OR NSCLC) AND (surg* OR resection OR lobectomy OR segmentectomy OR thoracoscop* OR perioperat*) AND (“pulmonary rehabilitation” OR prehabilitation OR telerehabilitation OR “respiratory training” OR “breathing exercise*” OR exercise) AND (“digital health” OR telehealth OR mHealth OR “mobile app*” OR smartphone* OR wearable* OR exergam* OR “virtual reality” OR avatar* OR chatbot* OR “conversational agent*” OR “digital human*” OR “digital therapeutic*” OR “artificial intelligence” OR “internet of things”)); PY=2024-2026
Scopus		TITLE-ABS-KEY((“lung cancer” OR “pulmonary neoplasm*” OR NSCLC) AND (surg* OR resection OR lobectomy OR segmentectomy OR thoracoscop* OR perioperat*) AND (“pulmonary rehabilitation” OR prehabilitation OR telerehabilitation OR “respiratory training” OR “breathing exercise*” OR exercise) AND (“digital health” OR telehealth OR mhealth OR “mobile app*” OR smartphone* OR wearable* OR exergam* OR “virtual reality” OR avatar* OR chatbot* OR “conversational agent*” OR “digital human*” OR “digital therapeutic*” OR “artificial intelligence” OR “internet of things”)) AND PUBYEAR > 2023 AND PUBYEAR < 2027
CINAHL		((MH “Lung Neoplasms+”) OR TI (“lung cancer” OR NSCLC) OR AB (“lung cancer” OR NSCLC)) AND (TI (surg* OR resection OR lobectomy OR thoracoscop* OR perioperat*) OR AB (surg* OR resection OR lobectomy OR thoracoscop* OR perioperat*)) AND ((MH “Pulmonary Rehabilitation”) OR TI (prehabilitation OR telerehabilitation OR “breathing exercise*” OR exercise) OR AB (prehabilitation OR telerehabilitation OR “breathing exercise*” OR exercise)) AND (TI (“digital health” OR telehealth OR mhealth OR “mobile app*” OR wearable* OR “virtual reality” OR avatar* OR chatbot* OR “digital human*” OR “artificial intelligence”) OR AB (“digital health” OR telehealth OR mhealth OR “mobile app*” OR wearable* OR “virtual reality” OR avatar* OR chatbot* OR “digital human*” OR “artificial intelligence”)); Published Date: 20240102-20260831
CENTRAL		([mh “Lung Neoplasms”] OR “lung cancer”:ti,ab,kw OR NSCLC:ti,ab,kw) AND (surg*:ti,ab,kw OR resection:ti,ab,kw OR lobectomy:ti,ab,kw OR thoracoscop*:ti,ab,kw OR perioperat*:ti,ab,kw) AND ([mh “Pulmonary Rehabilitation”] OR prehabilitation:ti,ab,kw OR telerehabilitation:ti,ab,kw OR “breathing exercise*”:ti,ab,kw OR exercise:ti,ab,kw) AND (“digital health”:ti,ab,kw OR telehealth:ti,ab,kw OR mhealth:ti,ab,kw OR “mobile app*”:ti,ab,kw OR wearable*:ti,ab,kw OR “virtual reality”:ti,ab,kw OR avatar*:ti,ab,kw OR chatbot*:ti,ab,kw OR “digital human*”:ti,ab,kw OR “artificial intelligence”:ti,ab,kw); custom date 2024-2026

Evidence architecture: Stage 1 retained the 11-study corpus of Lu et al.[4] and searched through January 1, 2024. Stage 2 was a targeted PubMed/MEDLINE and citation update from January 2, 2024, through August 31, 2026. The following strategies are provided to make the update reproducible and to permit the author team to perform a full multidatabase rerun. Retrieval, deduplication, and screening counts must be populated if those databases are rerun; they were not available in the submitted project archive

Eligibility criteria and study selection

Using the Population–Concept–Context framework, we included adults undergoing or recovering from lung cancer resection; digitally mediated pulmonary rehabilitation, breathing training, exercise, education, symptom monitoring, or self-management; and interventions delivered from the preoperative period through postoperative recovery. Telephone or messaging support was eligible only when embedded in a structured remote rehabilitation pathway. Purely diagnostic AI, surgical navigation, nonrehabilitation symptom monitoring, and studies without a surgical lung cancer population were excluded from the direct map. Evidence from conversational agent, digital biomarker, interoperability, and implementation research outside this population was retained only as contextual evidence for the translational framework.[12,13,14,15,16,17,18,19,20] The prespecified eligibility and data-charting framework is summarized in Supplementary Table 3.

Supplementary Table 3.

Eligibility and data-charting framework.

Item		Prespecified definition
Population		Adults scheduled for or recovering from lung-cancer resection
Concept		Digitally mediated pulmonary rehabilitation, breathing training, exercise, education, symptom monitoring, or self-management
Context		Preoperative period through postoperative recovery; hospital, home, or hybrid delivery
Direct-map exclusions		Diagnostic AI, surgical navigation, nonrehabilitation monitoring, nonsurgical or non-lung cancer cohorts, and contextual technology papers
Contextual evidence		Conversational agents, digital biomarkers, interoperability, implementation science, and AI reporting guidance used only to build the framework
Charted fields		Study design, sample, phase, components, sensing, adaptation, outcomes, implementation, safety, technology level, and risk-of-bias domains

AI=Artificial intelligence

The targeted update was screened by one reviewer. Included records and key effect estimates were verified against full reports or publisher/PubMed records. Data items included design, sample, perioperative phase, intervention components,[42] adaptation mechanism, outcomes, implementation findings, and technology level. No study authors were contacted.

Operational technology taxonomy

Technologies were classified using four prespecified dimensions: embodiment/interface, language generation, sensing, and decision logic [Table 1 and Figure 1]. Level 1 comprised fixed immersive or display systems, including VR, fixed video, animation, or exergaming, without patient-specific closed-loop adaptation. Level 2 comprised deterministic connected systems, including applications, wearables, scripted avatars, threshold alerts, dashboards, or clinician-mediated telemonitoring in which personalization followed preset rules, thresholds, or human input. Level 3 required all four of the following in deployed clinical use: an embodied conversational interface, generative natural language capability, multimodal sensing, and autonomous or adaptive decision logic. An “AI” label alone was insufficient. A hybrid system missing any Level 3 criterion was classified as Level 2; a protocol describing all criteria without completed outcomes was designated a Level 2/3 candidate.

Table 1.

Operational technology taxonomy and classification rules.

Dimension		Level 1: Fixed immersive/display		Level 2: Deterministic connected		Level 3: AI-driven digital human
Core mechanism		Fixed content, immersion, simulation, or demonstration		Preset rules, thresholds, dashboards, or clinician-mediated adaptation		Generative and context-sensitive interaction with autonomous/adaptive logic
Interface/embodiment		VR scene, fixed video, animation, exergame		App, wearable, scripted avatar, connected device		Embodied conversational interface
Language		None or prerecorded		Fixed scripts or menus		Generative natural language interaction
Sensing		Optional; no closed loop		Single or multiple sensors feeding rules/clinicians		Multimodal sensing integrated into adaptive decisions
Required decision rule		No patient-specific closed-loop adaptation		Personalization follows rules, thresholds, or human input		All four Level 3 capabilities documented in deployed use
Borderline rule		Fixed digital education remains Level 1		“AI” branding alone remains Level 2; hybrid systems missing one criterion remain Level 2		Protocol without completed outcomes=Level 2/3 candidate, not Level 3 evidence

Level 3 requires all four documented capabilities in deployed clinical use: embodiment, generative language, multimodal sensing, and adaptive decision logic. AI=Artificial intelligence

Figure 1.

Central illustration: Technology taxonomy and translational pathway. Solid borders denote completed direct clinical evidence; the dashed Level 3 border denotes a translational concept. Border style does not represent effect magnitude or certainty.

Methodological and implementation appraisal

For the inherited corpus, we retained the source review’s published methodological appraisal.[4] Updated randomized studies were examined using RoB 2 domains and nonrandomized studies using ROBINS-I.[7,8] Judgments are presented as provisional because duplicate assessment was not available. We did not apply Grading of Recommendations Assessment, Development, and Evaluation (GRADE): in a heterogeneous scoping map, evidence density and RoB should not be conflated with outcome-specific certainty.

Implementation findings were organized with the Nonadoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework; the extraction rules governing this distinction are provided in Supplementary Table 4.[9] To reduce interpretation bias, Table 2 distinguishes barriers and facilitators directly reported in primary studies from concerns inferred through the framework. Studies using AI terminology were additionally audited against relevant consolidated standards of reporting trials–artificial intelligence (CONSORT-AI) and standard protocol items: recommendations for interventional trials–artificial intelligence (SPIRIT-AI) reporting domains, including model identity/version, inputs, autonomy, human oversight, error handling, and monitoring.[10,11]

Supplementary Table 4.

Nonadoption, Abandonment, Scale-up, Spread, and Sustainability evidence extraction rules.

Category		Rule
Direct evidence		Statements explicitly reported by an included primary study or the source systematic review, including recruitment, usability, device/internet barriers, sensor concerns, management time, or provider data
Framework-derived concern		A plausible implementation issue identified by applying NASSS but not consistently measured in the mapped studies, including reimbursement, liability, EHR integration, cybersecurity, SaMD pathways, and sustainability
Reporting rule		Direct and inferred findings are displayed in separate columns. Inferred concerns are treated as hypotheses for future evaluation, not empirical results
Interpretive safeguard		Absence of reporting is not interpreted as absence of a barrier or facilitator

NASSS=Nonadoption, Abandonment, Scale-up, Spread, and Sustainability, SaMD=Software as a Medical Device, EHR=Electronic health record

Table 2.

Nonadoption, abandonment, scale-up, spread, and sustainability implementation synthesis separating reported evidence from framework-derived concerns.

NASSS domain		Directly reported in mapped studies		Framework-derived concern (not an extracted outcome)		Implication
Condition		Short perioperative window; older and functionally vulnerable population		Need for risk stratification across procedure type and comorbidity		Prespecify clinical exclusions and escalation rules
Technology		Device burden, internet/audio/video quality, sensor accuracy concerns[4]		EHR integration, cybersecurity, model drift, version control		Validate sensors and failure behavior before effectiveness testing
Value proposition		One trial reported lower management time[25]		Cost-effectiveness, opportunity cost, reimbursement		Prospectively collect costs, staff time, and utilization
Adopters		Recruitment 40%–70%; smartphone-capable samples; technology unfamiliarity[4]		Uncanny-valley response and trust in generative agents		Conduct age-, literacy-, and caregiver-specific usability studies
Organization		Provider experience largely unreported		Training, staffing, alert fatigue, accountability		Measure adoption, workload, and escalation burden
Wider system		Little direct evidence		SaMD classification, privacy, procurement, liability		Define intended use and governance before deployment
Embedding/adaptation		Short follow-up; maintenance rarely quantified		Long-term sustainability and model updates		Use RE-AIM and postdeployment monitoring

NASSS=Nonadoption, Abandonment, Scale-up, Spread, and Sustainability, SaMD=Software as a Medical Device, EHR=Electronic health record, RE-AIM=Reach, Effectiveness, Adoption, Implementation, Maintenance

Results

Evidence set and technology distribution

The inherited corpus contained 11 completed studies.[4,31,32,33,34,35,36,37,38,39,40,41] including fixed multimedia and animation education,[33,37,40] video-guided breathing training,[34] internet-based whole-course education,[33] connected applications and platform-based monitoring,[31,36,38,39] wearable or pedometer-based feedback,[32,41] and remote clinician- or messenger-supported programs.[35] The targeted update identified eight completed reports[21,22,23,24,25,26,27,28] and one ongoing protocol,[29] yielding 19 completed studies plus one protocol in the direct evidence map. Five completed interventions met Level 1 criteria (fixed multimedia, animation/video, or VR), 14 met Level 2 criteria, and none met Level 3 criteria. The Internet of things protocol included AI-enabled voice and sensor functions but had no completed outcomes and was retained as a Level 2/3 candidate.[29] The study-level direct evidence map is presented in Supplementary Table 5. Figure 1 displays the taxonomy and Figure 2 displays relative evidence density. Bubble size in Figure 2 indicates the number and breadth of reports only; it does not represent methodological quality, effect magnitude, or certainty.

Supplementary Table 5.

Direct clinical evidence map.

Study		Stage		Design/sample		Digital intervention		Phase		Level		Reason for classification
Kadiri et al., 2019[31]		Inherited		Quasi-experimental		Fit4Surgery app, pulse oximeter, biofeedback		Pre/post		2		Connected app/sensor feedback; deterministic
Finley et al., 2021[32]		Inherited		Pilot comparative		Exercise prescription plus activity tracker		Preoperative		2		Wearable monitoring with clinician prescription
Qin et al., 2023[33]		Inherited		Quasi-experimental		Internet-based whole-course education		Perioperative		1		Fixed digital education; no adaptive loop documented
Chu et al., 2020[34]		Inherited		Comparative		Video-mission breathing exercise		Perioperative		1		Fixed video demonstration
Liu and Pan, 2021[35]		Inherited		Comparative		Remote supportive exercise/self-efficacy support		Perioperative		2		Human-mediated remote personalization
Yang et al., 2022[36]		Inherited		Single-arm pilot		Smart after-care app plus monitoring devices		Postoperative		2		Connected monitoring and clinician follow-up
Li et al., 2021[37]		Inherited		RCT		Animation education program		Postoperative		1		Fixed animation; no adaptive loop
Sui et al., 2020[38]		Inherited		Comparative		WeChat education and rehabilitation		Postoperative		2		Messaging-based structured remote support
Ji et al., 2019[39]		Inherited		Prospective trial		mHealth pulmonary rehabilitation platform		Postoperative		2		Platform-based monitoring and management
Sun et al., 2017[40]		Inherited		RCT		Multimedia self-management intervention		Perioperative		1		Fixed multimedia education
Chen et al.[41]		Inherited		Comparative		Wearable pedometer-based rehabilitation		Perioperative		2		Connected wearable feedback
Liang et al., 2024[22]		Update		Quasi-experimental		Hybrid structured education/rehabilitation		Perioperative		2		Hybrid remote and clinician-supported program
Granger et al., 2024[23]		Update		RCT, n=116		Home exercise plus weekly telephone support		Postoperative		2		Deterministic program with clinician-mediated adaptation
Lv et al., 2025[24]		Update		RCT, n=136		App with symptom alerts, exercise, education		Postoperative		2		Threshold alerts; no generative agent
Xu et al., 2025[25]		Update		Noninferiority RCT, n=186 randomized		DTx-assisted education, exercise, symptom monitoring		Perioperative		2		Rule-based digital management; no Level 3 evidence
Li et al., 2025[26]		Update		RCT, n=47		App/wearable cardio-oncology telerehabilitation		Postoperative survivorship		2		Wearable/app prescription; AI label insufficient
Wenjie et al., 2025[21]		Update		Nonconcurrent controlled, n=190		VR education and breathing training		Immediate postoperative		1		Immersive/fixed VR; no adaptive agent
Shin et al., 2026[27]		Update		RCT, n=64		Messenger/telephone pulmonary telerehabilitation		Postoperative older adults		2		Human-mediated remote support
Chen et al., 2026[28]		Update		RCT, n=138		Intelligent breathing-training device		Postoperative		2		Connected device; no generative/embodied agent documented
Hang et al., 2026[29]		Update		Multicenter RCT protocol, planned n=340		IoT respiratory trainer with AI-enabled functions		Perioperative		2/3 candidate		Protocol suggests sensing/AI voice; no completed outcomes or full Level 3 validation

AI=Artificial intelligence, RCT=Randomized controlled trials, IoT=Internet of things, VR=Virtual reality

Figure 2.

Evidence-gap map showing relative density of direct clinical evidence by technology level and outcome domain. Bubble size indicates evidence density only, not methodological quality, effect magnitude, or certainty. Methodological context is shown separately.

Clinical outcomes

The most stable quantitative estimates remain those from the inherited meta-analysis.[4] Digital rehabilitation reduced the perioperative decline in 6-min walk distance by 15.00 m (95% confidence interval [CI] 4.34–25.65), was associated with fewer PPCs (risk ratio: 0.45, 95% CI: 0.30–0.66), and shortened chest tube duration by 1.53 days (95% CI: 0.12–2.95). Postoperative length of stay was not significantly reduced (mean difference: −1.42 days, 95% CI: −3.45 to 0.62; P = 0.17). These estimates arose from only two to four combinable studies and should not be generalized to AI-driven systems.

Updated evidence remained heterogeneous. A hybrid structured education program reported improvements in pulmonary and recovery outcomes but used a quasi-experimental design.[22] In a 116-participant assessor-blinded trial, a 3-month home program did not improve the primary self-reported physical function outcome; secondary outcomes favored the intervention for 6-min walk distance at 3 months (mean difference: 39.7 m, 95% CI: 6.8–72.6), global quality of life (7.1 points, 95% CI: 0.4–13.8), and exercise self-efficacy (16.0 points, 95% CI: 7.0–24.9).[23]

In the 136-participant POPPER trial, a smartphone application with symptom alerts, aerobic and respiratory training, and education improved the 30-day pulmonary function recovery ratio (79.32% vs. 75.73%; P = 0.040).[24] A 2025 noninferiority trial found digital therapeutic management noninferior to multidisciplinary management for forced expiratory volume in 1 s recovery (87.18% vs. 84.21%) and markedly reduced reported patient management time (1.48 vs. 16.67 min), without significant differences in length of stay or 90-day readmission.[25] A small wearable/app trial in postoperative survivors improved peak oxygen uptake and quality of life but was conducted 1 to 2 months after surgery and addressed cardio-oncology rehabilitation rather than immediate perioperative recovery.[26] Later trials of messenger-supported telerehabilitation in older adults and intelligent breathing devices reported functional or pulmonary improvements, but remained Level 2 systems.[27,28] The only completed VR-specific update was a nonconcurrent controlled study and therefore remained vulnerable to temporal confounding.[21] Table 3 summarizes the outcome evidence without converting density into certainty.

Table 3.

Summary of direct clinical outcome evidence.

Outcome/evidence source		Estimate or finding		Interpretation and limitations
PPCs; inherited meta-analysis[4]		Risk ratio: 0.45 (95% CI: 0.30 to 0.66)		Potential 55% relative reduction; only a small subset was combinable, and legacy trials had high risk of bias
Chest-tube duration; inherited meta-analysis[4]		Mean difference: −1.53 days (95% CI: −2.95 to −0.12)		Statistically significant pooled estimate; limited study count and heterogeneity
Postoperative length of stay; inherited meta-analysis[4]		Mean difference: −1.42 days (95% CI: −3.45 to 0.62; P=0.17)		Not statistically significant; should not be described as an established reduction
6-min walk distance; inherited meta-analysis[4]		15.00 m less perioperative decline (95% CI: 4.34 to 25.65)		Based on two combinable studies; clinical and methodological heterogeneity
Physical function; Granger et al.[23]		Primary self-reported outcome: mean difference: 1.0 point (95% CI: −6.0 to 8.0)		Primary outcome was null; secondary 6MWD, quality-of-life, and self-efficacy outcomes improved
Pulmonary function recovery; POPPER[24]		79.32% vs. 75.73%; P=0.040		Level 2 app with alerts, exercise, and education; single trial, not generative AI
FEV1 recovery/workflow; Xu et al.[25]		Noninferior FEV1 recovery; management time 1.48 versus 16.67 min		No significant length-of-stay or readmission difference; open-label single-center trial
Level 3 digital human		No completed direct trial; no effect estimate		No evidence of benefit for adherence, pulmonary function, PPCs, quality of life, or utilization

Estimates are not a GRADE assessment. CI=Confidence interval, FEV1=Forced expiratory volume in 1 s, PPC=Postoperative pulmonary complication, 6MWD=6-min walk distance, AI=Artificial intelligence, GRADE=Grading of Recommendations Assessment, Development, and Evaluation

Risk of bias and reporting quality

In the inherited evidence, all five randomized trials were rated at overall high RoB; among six quasi-experimental studies, one was rated high risk, and two used single-arm designs.[4] Updated randomized reports generally generated some concerns because of open-label delivery, single-center recruitment, attrition, or reliance on secondary outcomes. Updated nonrandomized studies were at serious risk of confounding. The detailed provisional appraisal is provided in Supplementary Table 6.

Supplementary Table 6.

Methodological quality and provisional risk-of-bias appraisal.

Evidence		Design		Appraisal		Key basis
Inherited corpus (5 RCTs)[4]		RCT		High risk overall		Inadequate reporting of randomization/concealment, blinding/contamination, missing data, or outcome measurement
Inherited corpus (6 quasi-experimental)[4]		Nonrandomized		1 high risk; 5 lower risk in source review		Two were single-arm; residual confounding remains relevant
Wenjie et al.[21]		Nonconcurrent controlled		Serious risk (provisional)		Temporal allocation and confounding
Liang et al.[22]		Quasi-experimental		Serious risk (provisional)		Nonrandomized design and confounding
Granger et al.[23]		Assessor-blinded RCT		Some concerns (provisional)		Unblinded intervention; attrition; favorable findings mainly secondary outcomes
Lv et al.[24]		Single-blinded RCT		Some concerns (provisional)		Single center and participant/provider unblinding; outcome procedures partly blinded
Xu et al.[25]		Open-label noninferiority RCT		Some concerns to high (provisional)		Single center; 186 randomized, 147 completed; noninferiority and attrition considerations
Li et al.[26]		Small RCT		Some concerns (provisional)		Small single-center sample; incomplete blinding
Shin et al.[27]		Small RCT		Some concerns (provisional)		Small sample and unblinded remote intervention
Chen et al.[28]		RCT		Some concerns (provisional)		Single center; reporting detail requires full duplicate verification
Hang et al.[29]		Protocol		Not applicable to outcomes		No completed outcome data

Legacy judgments are taken directly from Lu et al.[4] Updated-study judgments are conservative, report-level, single-reviewer appraisals and require duplicate author-team confirmation before submission. RCT=Randomized controlled trials

The AI-reporting audit was expanded beyond two selected examples to every direct study or protocol using AI, intelligent, or digital therapeutic terminology.[24,25,26,27,28,29] The item-level audit is presented in Supplementary Table 7. None of the completed systems documented the full Level 3 combination of generative language, multimodal sensing, adaptive autonomous decision logic, and embodied interaction. Model versions, failure modes, out-of-distribution behavior, and postdeployment monitoring were generally absent. Consequently, no field-wide claim about AI reporting quality or efficacy is made.

Supplementary Table 7.

Audit of artificial intelligence-labeled or artificial intelligence-adjacent direct studies.

Study		Model/version reported		Generative language		Multimodal sensing		Adaptive autonomous decisions		Human oversight		Final level
Lv et al.[24]		No deployable model described		No		App inputs/alerts		Threshold-based		Clinical alert pathway		2
Xu et al.[25]		No generative model/version		No		PRO and management data		Rule-based/clinician-supported		Multidisciplinary team		2
Li et al.[26]		AI-driven instructions claimed; model details absent		No documented		Wearable heart-rate data		Prescription logic; not autonomous generative		Clinical prescription		2
Chen et al.[28]		“Intelligent” device; model details limited		No documented		Breathing-device data		Feedback algorithm; not Level 3		Clinical rehabilitation context		2
Hang et al.[29]		Protocol mentions AI/voice; version not available		Potential, unverified		IoT respiratory sensing		Proposed; not outcome-validated		Protocol-based clinical management		2/3 candidate

No completed study met all four Level 3 criteria. Therefore, the review does not generalize reporting-quality judgments from two selected studies to the field. AI=Artificial intelligence, IoT=Internet of things

Implementation evidence

Directly reported recruitment rates in the inherited review ranged from 40% to 70%; technology unfamiliarity, device scarcity, time burden, Internet or audiovisual quality, and sensor accuracy were recurring barriers.[4] Most feasibility samples comprised people already able to use smartphones, limiting age-specific generalizability. Provider experience was rarely reported. One updated trial quantified potential workflow value through reduced management time,[25] but robust cost-effectiveness, staffing requirements, interoperability performance, and long-term maintenance were not evaluated.

The NASSS synthesis therefore separates observation from inference [Table 2]. Reimbursement, regulatory classification, enterprise integration, cybersecurity governance, and organizational sustainability are therefore important translational hypotheses, not outcomes consistently extracted from the mapped studies, and are presented separately as framework-derived concerns in Table 2.

Discussion

Principal findings

This update changes the interpretation of the field in three ways. First, digitally mediated rehabilitation has a growing clinical evidence base, but the evidence concerns fixed or deterministic systems rather than AI-driven digital humans. Second, the strongest pooled signal is for PPCs and chest tube duration, whereas the apparent reduction in length of stay is statistically uncertain. Third, recent Level 2 trials demonstrate plausible functional and workflow benefits, but their heterogeneity and methodological limitations preclude attributing a class effect to any specific technology.

Taxonomy as the principal contribution

The three-level taxonomy is intended to prevent category error. Level 1 increases exposure, immersion, or engagement; Level 2 creates deterministic feedback loops through sensors, thresholds, applications, and clinicians; Level 3 adds generative, multimodal, and adaptive agency. Because clinical risk increases with autonomy and context dependence, these levels are not merely labels: they imply different validation requirements. A scripted avatar should not inherit claims made for generative AI, and an application branded as AI should not be considered a digital human without documented Level 3 capabilities.

No completed perioperative lung cancer trial met Level 3 criteria. The proposed Level 3 architecture is therefore a translational hypothesis, not an evidence-based intervention. Adjacent literature on conversational agents supports feasibility and personalization concepts,[17,18,19,20] but cannot establish effects on PPCs, pulmonary function, quality of life, or utilization in this population.

Clinical integration, safety, and regulation

Digital systems map naturally onto ERAS and I COUGH components: fixed demonstration can support breathing instruction; connected systems can reinforce mobilization, capture symptoms, and trigger escalation; a future Level 3 system could add natural dialog, affect-aware coaching, and technique feedback.[2,3,30] However, the latter functions must be bounded by a human-in-the-loop model. High-risk recommendations should be rule-constrained, red-flag symptoms should trigger clinician escalation, and every generated recommendation should be auditable.

Interoperability is necessary for a digital intervention to become a clinical tool rather than a standalone gadget. Substitutable Medical Applications and Reusable Technologies (SMART) on fast healthcare interoperability resources (FHIR) can support electronic prescribing, structured return of patient-reported outcomes, and integration of validated digital biomarkers.[12,13] Before deployment, sensing accuracy, alert performance, missing data behavior, and subgroup calibration require external validation. Systems that provide patient-specific therapeutic guidance may also fall within jurisdiction-specific Software as a Medical Device pathways, requiring explicit intended use, risk management, change control, and postmarket monitoring.[43]

Ranked research priorities

A ranked agenda is presented in Table 4. Priority 1 is a pragmatic, adequately powered head-to-head trial comparing a fully specified Level 3 system with a best-in-class Level 2 comparator using a core outcome set. Priority 2 is formal AI safety evaluation and external validation, including hallucination testing, human override, subgroup calibration, and version control. Priority 3 is geriatric usability and digital exclusion research. Priority 4 is interoperability and operational performance. Priority 5 is health economic and workforce evaluation. Priority 6 is long-term real-world implementation using outcomes such as reach, adoption, fidelity, and maintenance.[14,15] A minimum core outcome and implementation set to operationalize these priorities is proposed in Supplementary Table 8.

Table 4.

Ranked research priorities for artificial intelligence-enabled perioperative pulmonary rehabilitation.

Rank		Research priority		Minimum design requirements		Decisive outcomes
1		Pragmatic Level 3 versus best Level 2 comparative trial		Adequate power; multicenter; prespecified core outcome set; transparent intervention specification		PPCs, validated function, adherence, quality of life, utilization
2		AI safety and external validation		Hallucination/red-team testing; human override; version control; external and prospective validation		Unsafe recommendation rate, false alerts, subgroup calibration, clinician override
3		Geriatric usability and equity		Older adults, low digital literacy, sensory/cognitive impairment, caregiver mode		Reach, task completion, acceptability, abandonment, disparities
4		Interoperability and operations		SMART on FHIR workflow; validated sensors; alert routing and downtime plans		Data completeness, latency, alert performance, workload
5		Health economics and workforce		Prospective microcosting and comparative economic analysis		Cost per outcome, staff time, downstream utilization, budget impact
6		Real-world implementation and sustainability		Hybrid effectiveness-implementation design; long follow-up; RE-AIM/implementation outcomes		Reach, adoption, fidelity, maintenance, safety over updates

PPC=Postoperative pulmonary complication, AI=Artificial intelligence, SMART=Substitutable Medical Applications and Reusable Technologies, FHIR=Fast Healthcare Interoperability Resources, RE-AIM=Reach, Effectiveness, Adoption, Implementation, Maintenance

Supplementary Table 8.

Recommended core outcome and implementation set.

Domain		Minimum outcomes
Clinical safety		Standardized PPC definition; adverse events; escalation events; readmission
Pulmonary recovery		FEV1 recovery trajectory; validated inspiratory muscle strength where relevant
Function		6MWD or cardiopulmonary exercise testing; physical-function patient-reported outcome
Behavior		Session completion, prescribed-dose adherence, sustained engagement, abandonment
Patient experience		Quality of life, anxiety/fatigue, usability, trust, caregiver burden
AI safety		Unsafe recommendation rate, hallucination rate, false alerts, missed red flags, clinician overrides
Equity		Age, education, digital literacy, sensory/cognitive impairment, rural access
Operations		Data completeness, latency, interoperability success, alert burden, staff time
Economics		Intervention cost, staff cost, downstream utilization, cost-effectiveness/budget impact
Implementation		Reach, adoption, fidelity, maintenance, sustainability

FEV1=Forced expiratory volume in 1 s, PPC=Postoperative pulmonary complication, 6MWD=6-min walk distance, AI=Artificial intelligence

Study limitations

This review has important limitations. The foundational corpus was inherited from a prior systematic review rather than reconstructed from all original database exports. The update was a targeted PubMed/citation search, was limited to English-language indexed reports, and used single-reviewer screening and provisional RoB appraisal; completeness and reproducibility are therefore lower than for a duplicated de novo systematic review. The direct map combines heterogeneous perioperative phases, technologies, and outcomes. Evidence-density categories are descriptive and are not a GRADE assessment. Some NASSS concerns were inferred and are labeled accordingly. Finally, no completed Level 3 study exists in the target population, so all claims about digital human benefit remain prospective.

Conclusion

Digitally delivered pulmonary rehabilitation can support perioperative lung cancer care, but the current evidence base belongs to fixed and deterministic technologies. There is no direct clinical evidence that AI-driven digital humans improve adherence, pulmonary function, PPCs, quality of life, or healthcare utilization. The proposed taxonomy and translational framework make this gap explicit and define the methodological, safety, interoperability, and implementation requirements that should precede clinical adoption.

Authors’ contributions

Yan Sun: Conceptualization, methodology, evidence synthesis, and writing – original draft. Xiu Jin: Project administration, evidence acquisition, and critical review. Xiaoyue Wei: Data curation and literature review. Zeya Wu and Pei Yang: Formal analysis and evidence synthesis. Minjie Ma: Clinical interpretation and writing – review and editing. Huaping Wei: Conceptualization, supervision, and writing – review and editing. Shangqing Xu contribution exactly as recorded in submission metadata. All authors reviewed and approved the final manuscript and are accountable for its content and integrity.

Conflicts of interest

There are no conflicts of interest.

Data availability statement

All data supporting this review are contained in the article and its Supplementary Materials. No individual participant data were generated.

Funding Statement

The study was funded by the Gansu Provincial Natural Science Foundation (23JRRA0934); Lanzhou Science and Technology Program (2025-2-155); Gansu Provincial Natural Science Foundation (23JRRA1597); and Provincial Science and Technology Program - 2024 Key R and D Program, International Cooperation Category (24YFWA011).