Mobile application-enabled lifestyle interventions for metabolic dysfunction-associated steatotic liver disease: a systematic review and meta-analysis

Abstract

Objectives

To evaluate the effectiveness, feasibility and acceptability of mobile application-enabled lifestyle interventions in adults with metabolic dysfunction-associated steatotic liver disease (MASLD).

Design

Systematic review and meta-analysis following Cochrane Handbook and Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidance.

Data sources

Medline (via PubMed), EMBASE, Scopus and ClinicalTrials.gov, inception to 12 August 2026, with reference-list and citation searching.

Eligibility criteria

English-language randomised controlled trials (RCTs) and single-arm prospective studies of interventions delivered primarily by mobile application in adults with, or at risk of, MASLD, reporting a liver-specific outcome (enzymes, steatosis or fibrosis).

Data extraction and synthesis

Two reviewers independently extracted data and assessed risk of bias (Cochrane Risk of Bias 2; National Heart, Lung and Blood Institute tool). RCT outcomes were pooled by random-effects meta-analysis and other outcomes synthesised narratively. Certainty was assessed using the Grading of Recommendations Assessment, Development and Evaluation.

Results

12 studies (6 RCTs, 6 single-arm studies; n=756) were included. There was no significant pooled effect on weight (k (number of studies)=6: mean difference −1.14 kg, 95% CI −2.93 to 0.65; I²=69%), alanine aminotransferase (ALT; k=5: −7.6 U/L, −20.8 to 5.6; I²=79.7%) or aspartate aminotransferase (k=4: −5.7 U/L, −12.5 to 1.0; I²=0%). Prespecified subgroup analysis showed ALT effect differed by intervention type (p=0.011; human-supported, k=4: −10.94 U/L, −26.45 to 4.57; automated, k=1: +2.87 U/L, −1.87 to 7.61), though neither was significant. Median retention was 83% (range 49%–95%).

Conclusions

Feasibility and retention signals are encouraging, but controlled evidence does not support consistent efficacy. Larger, longer trials with standardised endpoints, diverse populations and cost-effectiveness analyses are needed and should test whether human support is the active ingredient before recommendation alongside MASLD pharmacotherapies.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Metabolic dysfunction-associated steatotic liver disease (MASLD) affects approximately 30% of adults globally. Lifestyle modification is the major therapeutic focus and is now mandated alongside approved pharmacotherapies, including resmetirom and GLP-1 receptor agonists when used in people with liver fibrosis. Prior systematic reviews have evaluated broader digital behaviour change interventions in MASLD, including telemedicine, web-based and video platforms alongside mobile application-enabled lifestyle interventions and their reported effects have varied with the modalities and study designs included.

WHAT THIS STUDY ADDS

• To our knowledge, this is the first systematic review and meta-analysis focused exclusively on mobile applications as the primary vehicle of lifestyle intervention in MASLD (12 studies, n=756). The presence of a liver-specific outcome was predefined as a requirement, and we synthesised randomised and single-arm evidence separately. The meta-analysis of six randomised controlled trials demonstrated no statistically significant pooled effects on weight, liver enzymes or waist circumference, with high heterogeneity for weight, alanine aminotransferase (ALT) and waist circumference (I²=69%–87%). However, retention was generally high, indicating the interventions were feasible and acceptable, although engagement was reported inconsistently. Furthermore, a subgroup analysis found that human-supported interventions produced directionally greater ALT reductions than fully automated approaches (between-group p=0.011). Additional app-specific features that co-occurred with larger effects in individual studies, such as integrated self-monitoring and personalised feedback, warrant prospective testing.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Current controlled evidence is insufficient to recommend smartphone apps as efficacious standalone lifestyle interventions or as adjuncts to MASLD pharmacotherapy. However, this review identifies specific app features that warrant incorporation into future trials, which should be adequately powered, adopt standardised and clinically meaningful endpoints including fibrosis outcomes and prospectively test whether human coaching is the active ingredient distinguishing effective from ineffective app-based interventions.

Introduction

Metabolic dysfunction-associated steatotic liver disease (MASLD), formerly known as non-alcoholic fatty liver disease (NAFLD), is now recognised as the most prevalent chronic liver condition globally.¹ Affecting approximately 30% of the global adult population, MASLD encompasses a spectrum of hepatic pathology closely tied to metabolic syndrome, obesity, dyslipidaemia and type 2 diabetes mellitus.^1–3

Lifestyle intervention incorporating dietary change, physical activity and sustainable weight loss has been the cornerstone of MASLD management.⁴ Recent pharmacotherapy approvals of resmetirom for metabolic dysfunction-associated steatohepatitis with fibrosis, and GLP-1 receptor agonists such as semaglutide represent the first disease-modifying agents for this condition.^{5 6} However, both approvals explicitly require co-delivery with lifestyle intervention, reflecting the consensus that drug efficacy and durability depend on concurrent behavioural change.

Translating this into real-world success remains challenging. Less than 10% of individuals with MASLD achieve the desired weight loss, and over 75% of those who do lose significant weight regain the weight in the 3 years following cessation of lifestyle modification.⁷ Many patients struggle with adherence to lifestyle change and traditional programmes are often resource-intensive, requiring frequent face-to-face contact, multidisciplinary support and long-term follow-up, features not universally available in most healthcare systems. Attrition is a well-recognised challenge in weight loss interventions, with reported dropout rates of up to 80% depending on the nature of the programme.⁸ Given the rising global prevalence of MASLD, any lifestyle intervention must be scalable, readily integrated into existing models of care and deliverable in a time-effective and cost-effective manner.

Digital health platforms, including smartphone applications, text messaging, wearables and integrated digital therapeutics, offer a promising solution. By leveraging digital technology, such interventions can reach wider populations, enable real-time feedback and coaching and encourage self-management through personalised goal setting and tracking.⁹ Given the global prevalence of smartphones and the increasing consumer acceptance of digital health tools, the use of digital technology in MASLD has attracted considerable research interest.^{10 11}

Four systematic reviews have evaluated digitally delivered lifestyle intervention in MASLD, and each pooled heterogenous delivery modalities.^12–15 Albhaisi et al combined mobile health and web-based digital therapeutics across eight studies (n=1001), including three mobile application-enabled lifestyle interventions. This review reported overall effectiveness of digital interventions in regard to weight loss; however, the meta-analysis pooled within-group change from single-arm studies alongside randomised intervention arms.¹³ Kumar et al pooled seven randomised controlled trials (RCTs; n=621) of which four used a smartphone application and three delivered the intervention by text message or telephone. Significant reductions in weight, alanine aminotransferase (ALT) and aspartate aminotransferase (AST), though not body mass index (BMI), were reported.¹⁴ Zafar et al included only double-arm trials (11 trials, n=1521) and reported significant improvements in weight, BMI, ALT and AST, though not in waist circumference or fibrosis score; a subgroup analysis of four trials found greater weight loss with mobile application-enabled interventions (standardised mean difference (SMD) −0.59) than with other eHealth modalities (SMD −0.20), a difference the authors acknowledged was borderline.¹² Smith et al’s review applied promise analysis across eleven studies (n=1288) to identify the ‘active ingredients’ of digital behaviour change, again spanning applications, SMS, websites and video platforms. In each case, smartphone applications were analysed together with, or as a subset of, structurally different digital modalities.¹⁵

This systematic review focuses specifically on mobile applications as the primary vehicle for lifestyle intervention in MASLD. This focus enables a granular analysis of app-specific design features, user engagement strategies and practicalities of clinical implementation, which are considerations distinct from those applicable to telemedicine or web-based platforms. Crucially, by synthesising evidence from both RCTs and single-arm studies, this review aims to directly interrogate the disconnect between controlled and uncontrolled evidence, and to address the question of whether the consistently positive signals from uncontrolled studies are borne out in the more rigorous trial designs. This is of direct practical relevance to clinicians and policymakers considering whether to integrate app-based tools into hepatology care pathways as a mandated adjunct to pharmacotherapy.

Methods

Study design and search strategy

This systematic review and meta-analysis was conducted according to the Cochrane Handbook for Systematic Reviews and is reported following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (see online supplemental file 1 (research checklist) for the completed PRISMA checklist; figure 1).^{16 17}

Figure 1. PRISMA flow diagram. MASLD, metabolic dysfunction-associated steatotic liver disease; PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

Eligible articles were determined by a search in Medline (via PubMed, National Library of Medicine), EMBASE, Scopus and ClinicalTrials.gov for the period up to 12 August 2026. The search strategy used Boolean operators (AND, OR), parentheses, quotation marks and asterisks, with no limits applied. The search strategies used for all databases are presented in online supplemental material S1. Furthermore, manual searches of reference lists and citation tracking of included articles were also undertaken to identify additional relevant studies not captured by the electronic search. Duplicates were removed and accuracy was checked by one reviewer (CG). Titles and abstracts were evaluated by two independent authors (CG and MM), who worked independently. Full texts of potentially eligible studies were then retrieved and assessed in detail against the PICOS eligibility criteria. In case of disagreement, a third author (DF) assisted in the resolution.

Eligibility criteria (Population, Intervention, Comparison, Outcomes and Study Design)

Eligible studies were restricted to peer-reviewed articles published in English and involving human participants. The Population, Intervention, Comparison, Outcomes and Study Design (PICOS) strategy was used to identify inclusion criteria (online supplemental table S1).

Studies were eligible if participants had a diagnosis of MASLD without secondary causes or MASLD confirmed by imaging, transient elastography or biopsy. Studies recruiting from populations at risk of MASLD such as those with type 2 diabetes or obesity were also eligible. Interventions were eligible where a mobile application served as the primary vehicle for delivery of the lifestyle intervention. Interventions in which the application was adjunctive to another dominant delivery modality were excluded, including browser-based or web applications

Data extraction and quality assessment

Data were extracted independently by two reviewers (CG and MM), with discrepancies resolved by a third reviewer (DF). Extracted fields included patient characteristics, trial design, intervention components and all prespecified outcome measures. Quality assessment used the Cochrane RoB2 tool for RCTs and the National Heart, Lung and Blood Institute (NHLBI) Quality Assessment Tool for single-arm studies.

Grading the evidence

The certainty of evidence for each meta-analysed outcome (weight, ALT, AST and waist circumference) was assessed using the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) framework. As all pooled outcomes derived from RCTs, each began at high certainty and was downgraded one level for each domain rated serious across the five domains: risk of bias, inconsistency, indirectness, imprecision and publication bias. The criteria applied for downgrading are detailed in online supplemental table S3. Grading was performed independently by two reviewers (CG and MM), with discrepancies resolved by a third reviewer (DF).

Data synthesis

The Cochrane Handbook advises against combining randomised and non-randomised studies in a single meta-analysis, as outcomes measured under different designs vary systematically and add to heterogeneity.¹⁶ Accordingly, changes in weight, ALT, AST and waist circumference were pooled from RCTs only, using a random-effects model, while single-arm pre–post studies were synthesised narratively. Every RCT that reported extractable data for a given outcome entered the primary analysis for that outcome; a trial that lacked usable data for a particular outcome could not contribute to its pooled estimate and was described narratively instead. For weight, a sensitivity analysis excluding the trial at high risk of bias was performed to test the robustness of the pooled result.

Retention was defined as the proportion of enrolled participants completing follow-up. For randomised trials, intervention-arm completion was used; single-arm studies contributed their overall completion rate. Given studies varied in design, follow-up duration and how completion was defined, retention was summarised descriptively (median and range) rather than formally meta-analysed. Engagement metrics were not pooled due to substantial heterogeneity in definition and measurement across studies (ranging from percentage of days with valid device data to percentage of participants meeting a predefined interaction threshold) and are therefore summarised descriptively also.

Meta-analysis

Where studies reported medians and IQRs, the mean and SD were estimated using the method proposed by Wan et al, which approximates the mean as the average of the lower quartile, median and upper quartile, and estimates the SD by dividing the IQR by 1.35, assuming an approximately normal distribution of the data.¹⁸

For outcomes where only preintervention and postintervention means and SDs were reported, the mean change was calculated as the difference between the postintervention and preintervention means. The SD of the change was estimated using the preintervention and postintervention SDs and an assumed within-subject correlation (r).¹⁶ A correlation of r=0.9 was assumed for weight and r=0.6 for ALT and AST. Empirical pre–post correlations were back-calculated from Kwon et al, who reported baseline, follow-up and change score SDs. The estimated correlations exceeded 0.95 for weight and ranged from 0.64 to 0.78 for ALT and AST. The assumed correlations of 0.9 for weight and 0.6 for ALT and AST were therefore selected to provide conservative estimates of the variability of change.

Given heterogeneity in follow-up duration and control conditions, a random-effects model was used, with between-study variance estimated by restricted maximum likelihood. Given the small number of included studies, 95% CIs for the pooled effect were calculated using the Hartung-Knapp method. All analyses were conducted in R using the metan package (V4.5.3).¹⁹

Four outcomes were analysed: weight, ALT, AST and waist circumference. Fibrosis and steatosis measures and other non-invasive tests were not reported consistently across studies and were therefore synthesised narratively rather than meta-analysed. The pooled mean difference represents the difference between the intervention and control groups’ mean change from baseline. A negative mean difference indicates greater change in the intervention group than control. I² statistics are reported to quantify heterogeneity, the amount of variation in study results that is not due to chance. Values greater than 75% indicate high heterogeneity. Publication and reporting bias was to be assessed by visual inspection of funnel plots where a sufficient number of studies (≥10) contributed to a pooled outcome.

Subgroup analyses were performed for ALT, stratified by intervention type (automated vs human-supported), to explore sources of heterogeneity. For weight, the primary analysis included all six RCTs; a sensitivity analysis excluded the single trial at high risk of bias, which was the only outcome to which that trial contributed extractable data.

Results

Search results

A total of 538 articles were found from the three databases and clinicaltrials.gov website. After removing duplicates, 318 articles remained for screening of titles and abstracts and 258 articles were excluded. Full-text review was performed for the remaining 60 articles and 12 trials were identified that met the eligibility criteria. The manual reference check did not reveal any unidentified articles (figure 1).

Study characteristics

12 studies met the inclusion criteria comprising 756 patients. Six of the studies were RCTs and six were single-arm pre–post studies. The studies were published between 2020 and 2024. The trials ranged in duration from 4 to 48 weeks. The subjects’ mean age was 51 years (range: 42–60 years); 47% of the participants were male and 29% with type 2 diabetes mellitus. The subjects’ mean BMI was 33.3 kg/m² (range: 27.5–39.1 kg/m²). Eight of the interventions were delivered by mobile application-enabled lifestyle interventions alone, typically incorporating features such as dietary logging, step-count monitoring, weight tracking and automated behavioural feedback.^20–27 One application was delivered alongside a structured meal replacement product.²³ One intervention utilised an app and scheduled telephone calls, which provided coaching and personalised goal adjustment.²⁸ One intervention (evaluated in both a single-arm and RCT study) paired an app and email messaging to deliver educational content and reminders,^{29 30} and one intervention used an app alongside web-based educational videos.³¹ MASLD was confirmed in all participants in the six randomised trials. Among the single-arm studies, confirmation was by imaging or histology in four, by a validated steatosis prediction score in one (93% of participants) and was not defined in one (table 1).

Table 1. Summary of study and participant characteristics.

Study, country, design, duration (weeks)	Sample size (intervention vs control)	Age	Male (%)	BMI (kg/m2)	CAP (dB/m) and LSM (kPa)	MASLD diagnosis criteria	Comorbidities			Ethnicity
							Cirrhosis (%)	T2DM (%)	HTN (%)
Alkhouri et al26 (2024), Arizona, USA, single-arm study, 12	22	48	23	38.2	334 8.2	CAP >274 dB/m BMI >30	–	46	59	Caucasian 41% Hispanic 46% Black 5% Mixed 9%
Björnsdottir et al24 (2024), Iceland, single-arm feasibility, 12	38	60	39	37.6	344 6.4	CAP >294 dB/m and BMI >30 or metabolic syndrome or type 2 DM	0	50	71	Caucasian 100%
Cho et al23 (2024), South Korea, RCT, 4	60 (30 vs 30)	42	94	30.1	–	BMI >25, hepatic steatosis on US, ALT 40–200 IU/L	–	0	29	Not defined
Kaewdech et al25 (2024), Thailand, RCT, 24	122 (61 vs 61)	53	35	27.5	290 5.7	MASLD without secondary causes diagnosed at tertiary centre	0	16	29	Not defined
Kwon et al27 (2024), South Korea, RCT, 24	111 (51 vs 60)	49	62	29.8	–	MASLD without secondary causes diagnosed at tertiary centre	0	31	40	Not defined
Lim et al22 (2020), Singapore, RCT, 24	108 (55 vs 53)	46	63	30.5	–	Hepatic steatosis on imaging or histology	0	28	70	Asian 94% Other 6%
McKenzie et al31 (2021), Indiana, USA, single-arm pilot, 96	96	52	20	39.1	–	NAFLD-Liver Fat Score > −0.640 (93% participants)	–	0*	–	Caucasian 96% Black 4%
Sato et al21 (2023), Japan, single-arm, 48	20	52	53	32.0	–	NAS of ≥4	0	11	58	Not defined
Stine et al33 (2022), Pennsylvania, USA, RCT, 16	40 (20 vs 20)	52	28	36.2	–	Historical NAS of ≥4 or hepatic steatosis on imaging and FIB-4 ≥1.45 or LSM >8.2 kPa or FAST score >0.35	0	45	68	Caucasian 100%
Tincopa et al30 (2022), Michigan, USA, single-arm pilot, 24	40	53	53	33.9	329 7.2	Hepatic steatosis on imaging or histology	30	43	43	Caucasian 93% Otherwise not defined
Tincopa et al29 (2024), Michigan, USA, RCT, 24	70 (35 vs 35)	52	47	32.2	319 7.6	Hepatic steatosis on imaging or histology	33	50	–	Caucasian 83% Otherwise not defined
Zaharia et al28 (2021), Germany, single-arm feasibility, 12	29	56	42	32.6	326 –	Not defined	–	100	–	Not defined

ALT, alanine aminotransferase; BMI, body mass index; CAP, controlled attenuation parameter; FIB-4, Fibrosis-4 score; HTN, hypertension; LSM, liver stiffness measurement; MASLD, metabolic dysfunction-associated steatotic liver disease; NAFLD, non-alcoholic fatty liver disease; NAS, NAFLD Activity Score; RCT, randomised controlled trial; T2DM, type 2 diabetes mellitus.

Risk-of-bias assessment

The single-arm studies were assessed using the National Heart, Lung and Blood Institute (NHLBI) quality assessment tool for before-and-after studies with no control group. Four studies received a quality rating of good,^{21 24 26 28} with two studies being fair.^{30 31} This is because neither had a sample size power calculation and it was unclear as to whether all eligible candidates were enrolled. One of these studies also had a >20% attrition rate.³¹

The RCTs were assessed using the Risk of Bias 2 tool (RoB2). One study had a low risk of bias.²⁵ Four raised some concerns, most commonly in the domains of deviations from intended interventions, outcome measurement and selective reporting.^{20 22 23 27} One study was at high risk of bias, having been substantially disrupted by the COVID-19 pandemic with extensive missing data and limited power.²⁹ It reported extractable outcome data for weight only; it was therefore included in the primary weight meta-analysis and removed in a sensitivity analysis, and could not contribute to the ALT, AST or waist circumference analyses. The risk of bias assessment is available in online supplemental tables S2a,b.

Clinical outcomes

Clinical outcomes varied across studies; most reported weight loss, liver enzymes and liver fat or fibrosis markers. Some also assessed cardiometabolic factors (waist circumference, blood pressure, glycated haemoglobin (HbA1c)), relevant given that cardiovascular disease is the leading cause of mortality in MASLD.³² The main outcomes for each trial are listed in table 2. Primary outcomes included change in weight, liver enzymes, liver fat percentage and NAFLD activity score. Of the six RCTs, four reported a statistically significant benefit on their designated primary endpoint. These endpoints were not uniform, however, spanning weight, liver enzymes, liver steatosis and psychological outcomes.^{20 22 23 27} Six single-arm studies were also included; three were feasibility studies without an efficacy primary outcome^{24 30 31} and of the remaining three, all reported a significant improvement in their primary, although heterogenous, endpoints.^{21 26 28}

Table 2. Study outcomes.

Study	Outcomes
	Weight (kg)	Liver markers	Metabolic markers
Alkhouri et al26 (2024)	Mean % change −2.9, p=0.008	ALT mean % change −24.7, p=0.004 AST not reported CAP mean % change −5.7, p=0.019 LSM mean % change −0.94, p=0.404 MRI-PDFF % mean % change −13.2 p=0.007	None reported
Björnsdottir et al24 (2024)	Mean % change −3.2, p≤0.001	ALT (U/L) median absolute change 0.0 p=0.37 AST (U/L) median absolute change 0.4 p=0.53 CAP (dB/m) mean absolute change −33.3 p≤0.001 LSM (kPa) median absolute change 0.2 p=0.11 MRI-PDFF % mean % change −19.4 p≤0.001	Waist circumference (cm) mean absolute change −4.0, p≤0.001 HbA1c (mmol/mol) median absolute change 0.5 p=0.03 Systolic blood pressure (mm Hg) mean absolute change −6.0, p=0.18
Cho et al23 (2024)	Intervention: mean change −2.93; Control: mean change −2.74 Between-group p=0.762	ALT (U/L): Intervention: mean change −28.32; Control: mean change −10.67 Between-group p=0.006 AST (U/L): Intervention: mean change −13.28; Control: mean change −4.25 Between-group p=0.071	Waist circumference (cm): Intervention: mean change −2.73; Control: mean change −2.22 Between-group p=0.428 Systolic blood pressure (mm Hg): Intervention: mean change 0.64; Control: mean change −1.71 Between-group p=0.649
Kaewdech et al25 (2024)	Intervention: median change −1.3; Control: median change −1.0 Between-group p=0.599	ALT (U/L): Intervention: median change −4.0; Control: median change −9.0 Between-group p=0.189 CAP (dB/m): Intervention: mean change −19.9; Control: mean change −25.1 Between-group p=0.506 LSM (kPa): Intervention: median change −0.4; Control: median change −0.1 Between-group p=0.225	Waist circumference (cm): Intervention: median change −1.5; Control: median change −2.0 Between-group p=0.768
Kwon et al27 (2024)	Intervention: mean change −1.54; Control: mean change −0.79 Between-group p=0.26	ALT (U/L): Intervention: mean change −8.05; Control: mean change from baseline −5.73 Between-group p=0.62 AST (U/L): Intervention: mean change −4.09; Control: mean change −0.90 Between-group p=0.43	None reported
Lim et al22 (2020)	Intervention: mean change −3.2; Control: mean change −0.5 Between-group p≤0.001	ALT (U/L): Intervention: mean change −33.5; Control: mean change −11.5 Between-group p=0.004 AST (U/L): Intervention: mean change −17.4; Control: mean change −7.4 Between-group p=0.03	Waist circumference (cm): Intervention: mean change; Control: mean change 0.7 Between-group p≤0.001 Systolic blood pressure (mm Hg): Intervention: mean change −12.4; Control: mean change from baseline −2.4 Between-group p=0.003
McKenzie et al31 (2021)	Mean absolute change −7.3, p≤0.001	ALT (U/L) mean absolute change −5.4 p≤0.001 AST (U/L) mean absolute change −1.8 p=0.04	Waist circumference (cm) mean absolute change −8.0, p=0.002 HbA1c (mmol/mol) mean absolute change −2.2, p≤0.001 Systolic blood pressure (mm Hg) mean absolute change −2.6, p=0.18
Sato et al21 (2023)	Mean absolute change −12.4, p≤0.001	ALT (U/L) mean absolute change −49.53 p≤0.001 AST (U/L) mean absolute change −23.06 p=0.001 Mean NAS change −2.05±1.96 p=0.001	Waist circumference (cm) mean absolute change −4.1, p=0.009 HbA1c (mmol/mol) mean absolute change −0.27, p=0.2
Stine et al33 (2022)	Intervention: median change −8.5; Control: mean change +0.5 Between-group p=0.008	ALT (U/L): Intervention: mean change −9.2; Control: mean change −5.2 Between-group p=0.644 AST (U/L): Intervention: mean change −4.3; Control: mean change −3.9 Between-group p=0.602	None reported
Tincopa et al30 (2022)	Median absolute change 0.23, p=0.56	ALT (U/L) median absolute change −2.5 p=0.31 CAP (dB/m) mean absolute change −10.5 p=0.19 LSM (kPa) median absolute change 0, p=1	Waist circumference (cm) median absolute change 1.27, p=0.04 HbA1c (mmol/mol) median absolute change 0.1, p≤0.01 Systolic blood pressure (mm Hg) mean absolute change 3, p=0.16
Tincopa et al29 (2024)	Between group changes not reported Intervention: median weight loss 0 p=0.3; Control: median weight loss −0.25, p=0.4	CAP (dB/m): Intervention: median change −6 (n=1); Control: median change −49 p=0.65 Between-group changes not reported LSM (kPa): Intervention: median change 3.2 (n=1); Control: median change −1.95 p=0.17 Between-group changes not reported	None reported
Zaharia et al28 (2021)	Mean % change −10, p≤0.01	AST (U/L) mean % change −9 p=0.21 CAP (dB/m) mean % change −19, p≤0.01 LSM (kPa) mean absolute change −2.5, p≤0.05	None reported

Where relevant, primary outcomes are in bold.

ALT, alanine aminotransferase; AST, aspartate aminotransferase; CAP, controlled attenuation parameter; HbA1c, glycated haemoglobin; LSM, liver stiffness measurement; MRI-PDFF, MRI–proton density fat fraction; NAFLD, non-alcoholic fatty liver disease; NAS, NAFLD Activity Score.

Meta-analysis

Figure 2 displays the results of the meta-analysis. Six RCTs were included in the analysis.

Figure 2. Meta-analysis results for (a) weight (kg), (b) alanine aminotransferase, (c) aspartate aminotransferase and (d) waist circumference (cm). MD, mean difference.

Change in weight

When compared with standard of care, mobile application-enabled lifestyle interventions did not show a statistically significant change in weight (mean difference (MD) −1.14 kg, 95% CI −2.93 to 0.65, I²=69%) based on data from six RCTs (figure 2a). Removing the study with a high risk of bias that was disrupted by the COVID-19 pandemic produced similar results (MD −1.02 kg, 95% CI −2.90 to 0.77; I²=71%).

Change in liver-related outcomes

Alanine aminotransferase

When compared with standard of care, mobile application-enabled lifestyle interventions did not show a statistically significant change in ALT (MD −7.6 U/L, 95% CI −20.8 to +5.6, I²=79.7%) based on data from five RCTs (figure 2b).

A prespecified subgroup analysis of ALT by intervention type revealed a significant between-group difference (Q=6.45, df=1, p=0.011). The single automated intervention showed no significant reduction in ALT (MD +2.87 U/L, 95% CI −1.87 to +7.61; k=1), whereas the four human-supported interventions showed a directionally greater ALT reduction (MD −10.94 U/L, 95% CI −26.45 to +4.57; k=4). Neither subgroup estimate reached statistical significance individually, and the human-supported subgroup retained substantial heterogeneity (I²=65.4%). These findings should be interpreted with caution given the single study in the automated subgroup.

Aspartate aminotransferase

When compared with standard of care, mobile application-enabled lifestyle interventions did not show a statistically significant change in AST (MD −5.7 U/L, 95% CI −12.5 to 1.0, I²=0%) based on data from four RCTs (figure 2c).

Waist circumference

When compared with standard of care, mobile application-enabled lifestyle interventions did not show a statistically significant change in waist circumference (MD −1.1 cm, 95% CI −6.2 to +3.9, I²=86.9%) based on data from three RCTs (figure 2d).

Certainty of evidence

Using GRADE, the certainty of evidence was very low for weight, ALT and waist circumference and low for AST (online supplemental table S3). All four outcomes were downgraded for risk of bias and imprecision; weight, ALT and waist circumference were additionally downgraded for inconsistency owing to substantial unexplained heterogeneity, whereas AST was not (I²=0%). Publication bias could not be formally assessed given fewer than 10 trials per outcome.

Narrative synthesis of clinical outcomes

The pooled analyses showed no significant effect, from few trials with substantial heterogeneity. Single-arm findings must be treated cautiously: without a comparator, regression to the mean, placebo effects and self-selection mean positive signals cannot be taken as evidence of efficacy. The narrative synthesis below describes observed outcomes across designs to identify features worth testing in adequately powered trials.

Among the RCTs, interventions incorporating structured behavioural support and personalised feedback demonstrated the strongest signals of benefit. Lim et al reported that 40% of participants using the nBuddy app achieved >5% weight loss at 6 months, while Stine et al observed >5% weight loss in 45% of those using the Noom programme.^{20 22} By contrast, shorter or lower-intensity interventions did not demonstrate superiority over standard care.^{23 25} Five of six single-arm studies reported significant weight reduction, ranging from 2.9%²⁶ to 9%l²⁸, with Sato et al reporting histological benefit alongside an 8.3% weight reduction.²¹

Reductions in ALT accompanied weight loss in several studies.^{22 26 31} In Cho et al’s 4-week study, greater ALT reduction was observed in the intervention group despite the absence of significant weight change.²³

Steatosis and fibrosis measures dissociated in the two studies reporting the most complete non-invasive assessment.^{24 26} Alkhouri et al reported a 13.2% reduction in MRI–proton density fat fraction MRI-PDFF (p=0.007) alongside a non-significant 0.94% change in LSM (p=0.404), and Björnsdottir et al reported a 19.4% reduction in MRI-PDFF and a 33.3 dB/m fall in CAP (both p<0.001) with no change in LSM (+0.2 kPa, p=0.11).^{24 26} Fibrosis-related outcomes were assessed infrequently and none of the RCTs used histological endpoints. Sato et al were the only group to use histological endpoints, doing so in a single-arm study of 19 individuals without a control group, reporting fibrosis regression in over half of participants with F2–F3 disease.²¹ Kaewdech et al and Zaharia et al demonstrated statistically significant reductions in liver stiffness measurements.^{25 28}

Several studies reported improvements in cardiometabolic parameters including HbA1c, blood pressure and lipid profiles.^{24 28 30 31}

Digital therapeutics and app design

Application design varied widely, from text messaging and passive tracking to platforms regulated as medical devices (table 3). Interventions relying on unidirectional content delivery with minimal personalisation, such as the LINE-based programme of Kaewdech et al, demonstrated limited clinical benefit despite being low-cost and scalable.²⁵ Those incorporating interactive self-monitoring reported more favourable outcomes however this derives from comparison across heterogeneous studies rather than within-trial contrasts and cannot establish that these features drove benefit.^{21 22} Most interventions functioned as hybrid platforms pairing digital self-monitoring with human coaching or clinical oversight.^{20–24 27–31}

Table 3. Summary of intervention characteristics.

Study	Intervention	Control	Interventionist	Education	Dietary strategy	Physical activity monitoring	Interaction and feedback	Frequency
Alkhouri et al26 (2024)	CBT-based app delivering structured behaviour change modules for diet and exercise	–	Study staff for set up	⨯	Not defined	⨯	Algorithm	Weekly
Björnsdottir et al24 (2024)	Structured programme to improve diet, increase activity and reduce stress through behaviour change	–	Health coach	✔	Low carbohydrate	✔	Personalised health coach feedback	Daily
Cho et al23 (2024)	Dr Coach app designed to manage weight loss through nutritionist led coaching	Standard of care including brief education provided by healthcare providers	Nutritionist	✔	Partial meal replacement	✔	Personalised nutritionist feedback	Daily
Kaewdech et al25 (2024)	LINE app used to broadcast educational content about MASLD, diet and exercise	A single LSI video sent via app at enrolment	Study staff for set up	✔	Not defined	⨯	Nil	Weekly
Kwon et al27 (2024)	SMART-Liver app incorporating self-monitoring, coaching and education	Telephone calls at week 1, month 3 and month 6 with information on weight loss, diet and exercise	Dietician and physical activity specialist	✔	Hypocaloric	✔	Personalised dietician and physical activity specialist feedback	Monthly
Lim et al22 (2020)	nBuddy app with food diary, automated step count and weight tracking with dietician led coaching	Standard of care including brief education provided by a nurse	Dietician	✔	Culturally appropriate	✔	Personalised dietician feedback with monitored dashboard	Daily
McKenzie et al.31 (2021)	Support for nutrition and behaviour change, monitoring of biomarkers and medications through app about web videos	–	Health coach and clinician	✔	Ketogenic	⨯	Personalised health coach and clinician feedback	Daily
Sato et al.21 (2023)	NASH App to achieve lifestyle change via education, counselling and specific action goals	–	Clinician	✔	Not defined	✔	Personalised clinician feedback	Not defined
Stine et al33 (2022)	Noom app for weight loss through behaviour change	Dietary and exercise advice in line with international guidelines	Health coach	✔	Not defined	⨯	Personalised health coaching	Daily
Tincopa et al30 (2022)	Daily step count tracking via FitBit Zip watch and the FitBit app alongside hard copy educational material	–	Dietician	✔	Mediterranean	✔	Personalised dietician feedback	Weekly
Tincopa et al29 (2024)		Standard care with brief overview of lifestyle changes
Zaharia et al28 (2021)	Changing Health App consisting of an education programme, weight tracking and dietary monitoring	–	Nutritionist	✔	Hypocaloric	⨯	Personalised nutritionist telephone calls	Weekly

CBT, cognitive behavioural therapy; MASLD, metabolic dysfunction-associated steatotic liver disease.

Two interventions explicitly incorporated behavioural psychology: the cognitive behavioural therapy-based algorithm of Alkhouri et al²⁶ and the Noom programme.²⁰ Both reported significant weight loss, though neither design could separate the behavioural components from the rest of the intervention.

Integrated dietary and exercise strategies

Dietary and physical activity strategies varied considerably (table 3). Only one study used a Mediterranean diet despite guideline recommendations^{4 30}; others used hypocaloric, ketogenic or meal replacement approaches and physical activity monitoring typically relied on step-count tracking. Structured dietary interventions supported by coaching were associated with larger reductions in weight and steatosis markers, though largely from uncontrolled designs.^{21 22 28 31}

Engagement and retention data

Retention was among the more consistent findings across studies. Engagement, by contrast, was reported inconsistently and defined heterogeneously, which limited direct comparison (online supplemental table S4).

Retention was reported in all studies and was generally high (median 83%, range 49%–95%). The lowest value came from a randomised trial substantially disrupted by the COVID-19 pandemic, which explains the 49% floor.²⁹ Where reported, engagement varied widely and was not uniformly high; definitions ranged from the proportion of days with valid device data to the proportion of participants meeting a predefined interaction threshold. Several studies did not report engagement at all and some reported values below 75%. Higher engagement tended to occur in hybrid programmes combining automated support with human coaching, though this was not consistently measured.

Two studies reported within-study associations between engagement and outcome. Björnsdottir et al observed greater weight loss and liver fat reduction in more engaged users,²⁴ and Sato et al a correlation between engagement and histological improvement.²¹ In Tincopa et al participants who withdrew had worse baseline metabolic profiles.³⁰

Real-world feasibility and implementation

Several studies did not report on app acceptability and no studies performed a cost-effectiveness analysis. Feasibility-focused studies reported high satisfaction and completion rates.^{24 26 28 30} Enrolment rates, when reported, ranged from 62.5%³⁰ to 78%²⁰. Barriers included limited digital literacy, smartphone access and cohort homogeneity; several studies enrolled predominantly educated, single-ethnicity populations,^{20 24} and inability to use the application was cited as a reason for dropout in two trials.^{22 27}

Discussion

The growing burden of MASLD necessitates scalable, sustainable strategies to deliver lifestyle intervention. This review demonstrates that mobile application-enabled lifestyle interventions are consistently feasible and widely acceptable to participants, with generally high retention across studies, particularly in programmes combining automated support with human coaching. No pooled estimate reached statistical significance, and all four CIs were wide and spanned the null. Although several individual trials reported statistically significant effects on their own primary endpoints, these used different outcomes, follow-up durations and sample sizes, and most carried some concern or high risk of bias. This might be read as evidence of ineffectiveness, but the pooled estimates for weight, ALT and waist circumference carried very high heterogeneity (I²=69%–87%), were drawn from few trials varying widely in design, intensity and duration, and were generally underpowered to detect the modest effects anticipated for lifestyle interventions. These are the same concerns that placed the certainty of evidence at low (AST) to very low (weight, ALT and waist circumference) on GRADE. The absence of a demonstrable pooled effect more plausibly reflects the limited quantity and quality of the current evidence than established inefficacy.

The most directly comparable synthesis supports rather than contradicts our findings. Smith et al, analysing digital behaviour change interventions in MASLD as MDs, reported no significant effect on weight (WMD −2.07 kg, 95% CI −6.08 to 1.94), ALT (−9.14 IU/L, −20.33 to 2.05) or AST (−5.81 IU/L, −12.96 to 1.35), estimates closely concordant with ours despite independent conduct and different eligibility criteria.¹⁵

Our estimates require reconciliation with the significant effects reported by Kumar et al and Zafar et al.^{12 14} The differences are largely compositional and methodological. Both pooled interventions delivered by telephone, text message or web platform alongside applications; three of Kumar’s seven trials were not app-delivered, and Zafar’s app subgroup comprised four trials, three of which we also include, the fourth being a supervised exercise-training trial in 24 participants that contributed the largest effect in that subgroup.³³ Both also derived pooled effects by methods we did not adopt: Zafar et al pooled SMDs of change scores using an imputed within-subject correlation and, where arm-level counts were unavailable, an assumed 1:1 allocation, while Kumar et al’s weight analysis combined end-of-treatment and change-from-baseline values, with a single 28-participant feasibility trial contributing over 80% of the pooled weight. Both searches also closed in 2024 and predate two of our included RCTs. Notably, Zafar et al’s waist circumference and fibrosis score analyses were null, as were ours.

Uncontrolled evidence continues to appear more favourable. Avery et al’s single-arm feasibility study, which fell outside our eligibility criteria as the intervention was delivered via a web-app rather than a mobile application, reported 4.0 kg weight loss and a 27% ALT reduction in 35 patients.³⁴ AST, liver stiffness, FIB-4 and NAFLD Fibrosis Score all crossed the line of no effect and engagement declined sharply despite high initial uptake. This mirrors the pattern within our own single-arm studies, which likewise reported consistent weight and enzyme improvements without controlled corroboration and suggests the disconnect between controlled and uncontrolled evidence is a feature of study design rather than of any particular delivery modality.

Beyond these statistical considerations, the outcome of greatest clinical importance was structurally unreachable in the current trial base. Fibrosis is the strongest determinant of liver-related and all-cause mortality in MASLD, and simple steatosis is more appropriately regarded as a risk state than as the outcome of clinical interest.³⁵ Fibrosis outcomes were eligible for inclusion in this review but could not be meta-analysed. No randomised trial employed histological endpoints; liver stiffness measurement was reported in a minority of studies and inconsistently, as median or mean absolute change without dispersion measures suitable for pooling. FIB-4 and NAFLD Fibrosis Score were more consistently reported but are risk stratification tools, not validated for within-patient change. Fundamentally, the magnitude of change required was not achieved. Paired biopsy studies indicate that fibrosis improvement requires weight loss of approximately 7%–10%, and Luo et al found that even under a standardised 48-week lifestyle programme, fibrosis regression required relative weight loss of 4.94% in patients with elevated ALT and 8.55% in those with persistently normal ALT.³⁶ Our pooled estimate was −1.14 kg, and only the upper end of the single-arm range (2.9%–9%) approaches these thresholds. Short-term liver stiffness is also confounded by concurrent changes in steatosis, inflammation and body weight. The absence of a fibrosis endpoint therefore reflects the immaturity of the trial evidence rather than a design choice of this review and reinforces the need for adequately powered trials of sufficient duration and intervention intensity to achieve the weight loss at which fibrosis becomes detectable.

The one source of heterogeneity we were able to interrogate directly was intervention type. The subgroup difference in ALT was significant (p=0.011), with greater reductions in human-supported than in automated interventions and aligns with narrative evidence across the included trials that clinician or coach input accompanies stronger hepatic signals. Interventions incorporating structured behavioural support and personalised feedback typically integrated multiple behaviour change techniques, including self-monitoring, goal setting, feedback loops and reinforcement. Narratively, this pattern extended to weight and metabolic outcomes, although the ALT difference was the only one supported meta-analytically and the pooled weight comparison remained null.

The same pattern is visible in the wider digital health literature, where app-based interventions have shown clinically meaningful effects as the evidence base matured and heterogeneity has often proven explicable rather than intrinsic. In obesity, the most mechanistically comparable field, a 2025 umbrella review (261 studies; >62 000 participants) found a significant weight reduction (MD −1.32 kg), with heterogeneity (I²=86%) comparable to ours. Notably, interventions incorporating human support showed a consistent effect with no heterogeneity (I²=0%), whereas those without showed no effect and extreme heterogeneity (I²=96.5%).³⁷ Similarly, a recent BMJ Open meta-analysis of multimodal mobile application-enabled lifestyle interventions in diabetes found pooled averages masked substantial design-dependent variation, concluding that intervention design is a key determinant of effectiveness.³⁸

Two caveats temper this. One arm of our subgroup analysis rested on a single automated trial. Separately, Avery et al’s comparison of coaching versus self-regulation tools alone included 17 participants, of whom only eight attended any coaching appointment.³⁴ Whether human support is the active ingredient therefore remains a hypothesis for prospective testing rather than an established effect.

ALT itself may contribute further heterogeneity. Luo et al reported an approximately threefold difference in baseline ALT across MASLD phenotypes within a single cohort, with around a third of patients with fibrosis having persistently normal values.³⁶ Absolute ALT change is therefore constrained by cohort composition before any intervention is delivered, and trials enrolling different proportions of these phenotypes will yield divergent absolute effects for identical interventions. The same cohort found that ALT response did not differ across MASLD categories while liver stiffness response did, indicating that ALT change tracks the outcome of clinical interest imperfectly. Standardised reporting of percentage change from baseline or of the proportion achieving a predefined ALT response threshold, would improve comparability in future trials.

Sex is a further, largely unexamined source of heterogeneity. MASLD is sexually dimorphic, with higher prevalence in men until midlife and greater risk of advanced fibrosis in women after menopause.^{39 40} Sex may also modify adherence to lifestyle intervention, in which men are under-represented.⁴¹ Male participation ranged from 28% to 94% across the included RCTs and a mean age of 51 years places many women within the menopausal transition, yet no study reported outcomes by sex or recorded menopausal status. Two single-arm studies found no sex difference in retention or engagement but were underpowered.^{24 30} Meta-regression was not feasible with five to six trials per outcome and would have been confounded; the most male-dominated trial was also the shortest and among the most intensive.

Retention was generally high across studies, a meaningful feasibility signal, though it should be read against wide variation in follow-up duration and completion definitions. Engagement was reported too inconsistently to characterise reliably; where higher engagement was reported it was often associated with better outcomes, but this association is confounded by participant motivation and cannot be interpreted causally. Barriers such as digital literacy and device access were acknowledged but not insurmountable; interventions using onboarding support or widely available technologies (eg, LINE, Fitbit, Noom) expanded reach with minimal infrastructure. Nevertheless, many trials enrolled motivated, digitally literate participants, potentially limiting generalisability and overestimating real-world effectiveness.

Several studies demonstrated clinician use of patient progress dashboards supporting in-clinic counselling and potential hybrid care models.^{21 27} Standardisation of digital reporting is an unmet need with few studies using validated tools such as the Mobile App Rating Scale or collected granular engagement metrics, limiting identification of which design features drive outcomes.

This review has several strengths. To our knowledge it is the first synthesis focused exclusively on mobile application-enabled lifestyle interventions in MASLD; it was conducted with independent duplicate screening, data extraction, risk-of-bias assessment and GRADE rating; and it deliberately separated randomised from single-arm evidence, pooling only the former and treating uncontrolled data narratively. Several methodological limitations should nonetheless be acknowledged. The search was restricted to a limited number of databases and to peer-reviewed, English-language studies, raising the possibility of language and publication bias; with fewer than 10 trials contributing to any outcome, publication bias could not be formally assessed. Restricting eligibility to interventions in which a mobile application was the primary delivery vehicle improves interpretability by removing structurally dissimilar modalities but narrows comparability with broader digital-health reviews and excludes web-delivered MASLD interventions. The small number of trials per outcome also precluded meta-regression, so heterogeneity could be explored only through a single prespecified subgroup analysis, one arm of which rested on a single automated trial. Pooling required distributional assumptions: estimation of means and SD from medians and IQRs, and imputation of pre–post correlations (r=0.9 for weight, r=0.6 for aminotransferases), which may have introduced error into the pooled estimates, although the imputed correlations were chosen conservatively. Marked heterogeneity in outcome definitions, endpoints and engagement metrics further limited comparability and pooling. Finally, the included trials were themselves small, short in duration, mostly at some concern or high risk of bias, and enrolled predominantly motivated, digitally literate and ethnically homogeneous cohorts, with variable sex composition and no sex-stratified reporting, which together account for the low-to-very-low certainty of evidence and limit generalisability.

A more fundamental gap concerns the position these tools will occupy in future care. None of the included studies evaluated apps as adjuncts to pharmacotherapy, and none alongside GLP-1 receptor agonist treatment. As pharmacotherapy becomes established in MASLD, the role of digital tools may shift towards supporting patients on concurrent treatment. Future trials should therefore consider outcomes beyond direct liver-related efficacy, including medication adherence, side-effect management and sustained engagement.

Conclusions

Our findings indicate that, while mobile application-enabled lifestyle interventions show promise as an acceptable and scalable delivery platform, the current evidence base does not support their routine clinical implementation as efficacious tools. The priority for the field is larger, adequately powered RCTs of sufficient duration to detect changes in clinically meaningful endpoints, including liver histology and fibrosis-related outcomes. Such trials should prespecify and execute subgroup analyses by intervention type, intensity and duration, collect validated digital engagement metrics, report retention and feasibility alongside clinical outcomes, and recruit from ethnically and socioeconomically diverse populations while addressing digital access barriers. Cost-effectiveness analyses accounting for clinician time and resource utilisation are also essential for policy-level decision-making. Encouraging engagement and feasibility data should motivate rigorous prospective evaluation rather than premature clinical adoption.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.