Efficacy and safety of platelet-rich plasma injections for the treatment of knee osteoarthritis: a systematic review and meta-analysis of randomized controlled trials

Abstract

Introduction

Knee osteoarthritis (KOA) is a prevalent degenerative joint disorder affecting a significant portion of the elderly population. Despite the availability of various non-surgical and pharmacological treatments, their effectiveness is often limited by temporary symptom relief and lack of disease-modifying properties. Platelet-rich plasma (PRP) has emerged as a promising biological therapy for KOA, with preclinical evidence suggesting its potential to promote cartilage repair and modulate inflammation. This systematic review and meta-analysis aims to comprehensively evaluate the efficacy and safety of PRP injections in the treatment of KOA.

Methods

A systematic literature search was conducted from January 1, 2021, to December 31, 2024, encompassing major medical databases, clinical trial registries, and grey literature sources. Randomized controlled trials (RCTs) comparing PRP with other treatments for KOA were included based on predefined eligibility criteria. Data extraction and analysis were performed using various statistical software packages and machine learning models. A neural network model was constructed to predict PRP treatment outcomes by integrating multidimensional clinical features. Study quality was assessed using the Cochrane Risk of Bias Tool, and publication bias was evaluated through funnel plot analysis and Egger's test.

Results

The meta-analysis included 28 RCTs with a total of 3246 KOA patients. PRP demonstrated comparable pain relief to hyaluronic acid (HA) but superior functional improvement, especially when combined with HA. Compared to corticosteroids, PRP showed no significant difference in efficacy as monotherapy but enhanced outcomes when used in combination. PRP also outperformed physical therapy and exercise therapy in both pain control and functional improvement. The optimal PRP concentration range was identified as 600–900 × 10⁹/L, with 3–5 injections at 7–14-day intervals yielding the best results. Early intervention, particularly in KL grade I–II patients, was associated with superior outcomes. The neural network model accurately predicted treatment responses based on patient characteristics and disease factors.

Discussion

The findings of this study have important implications for understanding the individualized regulatory mechanisms of PRP therapy. The nonlinear relationship between PRP concentration and treatment efficacy reflects the complex cytokine network dynamics and receptor saturation effects. The superiority of the 3–5 injection regimen may be attributed to its alignment with the time window of chondrocyte gene expression regulation, potentially mediated by epigenetic mechanisms. The synergistic effects of PRP with HA and the time-dependent treatment response patterns provide new insights for developing personalized, multi-target treatment strategies. The deep learning model demonstrated the potential of data-driven appjjroaches for optimizing PRP therapy and highlighted the need to address challenges in biological interpretability, data standardization, and clinical implementation.

Conclusions

This comprehensive evaluation of PRP therapy for KOA identified key parameters associated with optimal treatment outcomes, including PRP concentration, injection frequency, and early intervention. PRP demonstrated unique value and synergistic effects when combined with other treatments, and machine learning models provided new avenues for personalized treatment optimization. Future research should focus on addressing limitations in long-term follow-up data, standardizing evaluation criteria, and exploring the clinical application of artificial intelligence techniques to enhance the precision and effectiveness of PRP therapy for KOA.

Introduction

Knee osteoarthritis (OA) is a prevalent chronic degenerative joint disease primarily affecting middle-aged and elderly populations. Globally, approximately 13% of women and 10% of men over 60 years of age suffer from symptomatic knee OA, with incidence rates significantly increasing with age [1]. The condition manifests through various clinical presentations, including joint pain, limited mobility, swelling, morning stiffness, and deformity [2]. Beyond affecting physical functionality, knee OA can lead to psychological complications, such as depression and anxiety, substantially diminishing patients' quality of life [3].

Current conventional treatments for knee OA primarily comprise non-surgical and pharmacological interventions. Non-surgical approaches encompass physical therapy modalities (including heat therapy, electrical stimulation, and strength training), exercise therapy (such as low-impact activities and joint flexibility training), and lifestyle modifications (including weight management and gait correction) [1]. In terms of pharmacological management, while non-steroidal anti-inflammatory drugs (NSAIDs), analgesics, and hyaluronic acid (HA) injections are widely utilized, long-term NSAID use poses risks of gastrointestinal and cardiovascular adverse effects, and HA injection efficacy shows significant individual variation [4].

Despite providing some degree of pain relief and functional improvement, traditional treatments face numerous limitations. Most notably, these interventions offer only temporary relief without halting disease progression [5]. Furthermore, poor patient compliance with oral medications or injection therapies often compromises treatment effectiveness. Although physical therapy and exercise interventions are widely recommended, many patients struggle with long-term adherence and treatment outcomes vary considerably based on individual factors [1]. When conservative treatment proves inadequate, surgical intervention becomes necessary for some patients. However, while surgical procedures such as joint replacement can significantly improve joint function, not all patients are suitable candidates, and surgical risks, including infection and post-operative complications, cannot be overlooked [6]. Consequently, the current challenge lies in identifying non-surgical treatment options that can effectively control symptoms while potentially slowing disease progression, improving patient quality of life, and reducing surgical dependency [2].

Platelet-rich plasma (PRP) therapy has emerged as a promising treatment for knee OA, demonstrating potential through multiple mechanisms, including cartilage repair promotion, inflammation modulation, and joint microenvironment optimization. PRP contains concentrated levels of various growth factors, including platelet-derived growth factor (PDGF), insulin-like growth factor (IGF-1), vascular endothelial growth factor (VEGF), and transforming growth factor-β (TGF-β), which promote chondrocyte proliferation, enhance extracellular matrix (ECM) synthesis, and inhibit cartilage degradation, potentially slowing joint degeneration [7]. Moreover, PRP can maintain joint homeostasis by suppressing pro-inflammatory factors (IL-1β and TNF-α), reducing synovial inflammation, improving synovial fluid composition, and enhancing hyaluronic acid synthesis [8]. Animal studies have further confirmed that intra-articular PRP injection can reduce inflammation levels, alleviate cartilage damage, and potentially exert disease-modifying effects [9].

Recent clinical studies and meta-analyses have validated PRP's therapeutic efficacy in knee OA treatment. Compared to saline and corticosteroids, PRP demonstrates superior efficacy in pain relief and functional improvement [10], particularly in early to moderate-stage knee OA (Kellgren–Lawrence grades 1–3) patients [11]. In addition, PRP combined with hyaluronic acid (HA) has shown promising trends at specific timepoints, especially regarding improved safety profiles [12]. While consensus remains elusive regarding optimal PRP preparation methods, injection frequency, and comparisons between leukocyte-poor and leukocyte-rich PRP, this biological therapy shows considerable promise. Further optimization of treatment protocols and validation of long-term efficacy and safety through large-scale, high-quality randomized controlled trials (RCTs) remain essential for future research [13].

Methods

Literature search strategy

This study employed a systematic literature search methodology from January 1, 2021, to December 31, 2024. This time frame was selected based on several key considerations: primarily, the rapid advancement of PRP treatment technology in recent years, ensuring that studies within this period better reflect current clinical practices; second, studies conducted after 2021 generally adopted more standardized methodological criteria, contributing to higher quality evidence; and finally, research during this period increasingly focused on optimizing personalized treatment protocols, closely aligning with the objectives of this study.

The search process encompassed major medical literature databases, including English-language databases (PubMed, Web of Science Core Collection, Cochrane Library, EMBASE), clinical trial registries (ClinicalTrials.gov, WHO ICTRP, ChiCTR), and grey literature databases (OpenGrey, ProQuest Dissertations, and Theses). The Cochrane Library search specifically included three sub-databases: CENTRAL, CDSR, and DARE.

The search strategy integrated both Medical Subject Headings (MeSH) terms and free-text keywords. The primary search terms were categorized as follows:

• Treatment-related terms: ("Platelet-Rich Plasma"[Mesh] OR "PRP"[Title/Abstract] OR "Platelet Rich Plasma"[Title/Abstract])

• Disease-related terms: ("Knee Osteoarthritis"[Mesh] OR "Knee OA"[Title/Abstract] OR "Osteoarthritis, Knee"[Title/Abstract])

• Study type-related terms: ("Randomized Controlled Trial"[Publication Type] OR "Clinical Trial"[Publication Type])

To ensure comprehensive coverage, the research team implemented additional measures: first, manually reviewing reference lists of all included studies; second, contacting leading researchers in the field to identify unpublished or ongoing studies; and finally, searching relevant conference proceedings to capture potentially overlooked research. Two researchers independently conducted all search processes experienced in systematic reviews, with any discrepancies resolved through discussion with a senior researcher. The initial search yielded 107 relevant articles.

Inclusion criteria

Study design: Only randomized controlled trials (RCTs) with clearly described randomization methods and established control groups were included. Following the PRISMA flow diagram, nine non-RCT studies were excluded.

Study population: The study population comprised clinically and radiologically confirmed knee osteoarthritis patients. Eligible patients were required to present definitive clinical symptoms and signs, with radiological examinations (X-ray, MRI) meeting established diagnostic criteria for osteoarthritis. Studies were required to report baseline patient characteristics, including age, gender distribution, and disease duration.

Intervention protocols: For the experimental groups receiving PRP treatment, studies were required to document explicitly:

1. PRP preparation methods and quality control measures

2. PRP concentration and activity indicators

3. Detailed injection protocols, including frequency, intervals between injections, and follow-up duration

4. Control group interventions (e.g., placebo, hyaluronic acid)

The inclusion of studies with various control interventions was intentional. While our primary focus is evaluating PRP therapy, comparing PRP with other established treatments (hyaluronic acid, corticosteroids, physical therapy) provides essential clinical context. This comparative approach enables more meaningful interpretation of PRP's effectiveness, identifies potential synergistic effects in combination therapies, and offers evidence-based guidance for treatment selection in real-world clinical settings.

As illustrated in Fig. 1, 12 studies lacking control groups were excluded.

Fig. 1.

Study selection process. PRISMA flow diagram showing the systematic review process. From 107 initial articles, 28 studies met inclusion criteria

Outcome measures: Primary outcomes included standardized pain assessments (e.g., VAS scores), functional evaluation scales (e.g., WOMAC scores), and quality-of-life assessments. Secondary outcomes encompassed adverse event rates and patient satisfaction measures. Studies were required to provide complete follow-up data and outcome evaluations. The flow diagram indicates that eight studies with incomplete data were excluded.

Following rigorous screening and quality assessment, 28 studies meeting the criteria above were ultimately included for qualitative analysis. These studies provided comprehensive methodological information and resulted in data adequately supporting subsequent data extraction and analysis.

Two researchers with expertise in systematic reviews independently conducted all literature searches, study selection, and data extraction. To ensure consistency, both researchers underwent calibration exercises using a pilot sample of 10 studies before beginning the formal review process.

Discrepancies between reviewers were resolved through the following process:

• Initial disagreements were discussed in detail with reference to the original inclusion/exclusion criteria

• For data extraction discrepancies, both reviewers re-examined the source documents together

• In cases where consensus could not be reached (estimated < 5% of decisions), the study team held a consensus meeting to reach a final decision

• A detailed record of all disagreements and their resolutions was maintained to ensure transparency

While a third independent reviewer was not employed due to resource constraints, we implemented multiple quality control measures, including: (1) use of standardized extraction forms, (2) regular team meetings to discuss challenging cases, and (3) random verification of 20% of extracted data by cross-checking between reviewers. These measures ensured the reliability and reproducibility of our review process.

Data extraction and analysis

Data preprocessing methods

This study utilized multiple statistical software packages and analytical methods for data processing. Meta-analysis was conducted using Review Manager (RevMan) version 5.4.1, machine learning models were constructed using Python 3.9.0 environment, statistical analyses were performed using R 4.2.1 (including meta package version 6.2–1), and sensitivity analyses were completed using STATA 17.0. Researchers with statistical expertise conducted all analytical procedures and underwent cross validation (Table 1).

Table 1.

Systematic review of PRP treatment for knee osteoarthritis: data summary

Study (year)	PRP characteristics and preparation	Treatment protocol	Clinical outcomes	Adverse events
Romandini et al. [14] (2024)	LR-PRP: 1009.3 ± 213.9 × 10⁹/L, WBC 14.4 ± 4.9 × 109/L LP-PRP: 939.3 ± 190.3 × 10⁹/L, WBC 4.7 ± 2.4 × 10⁶/L Preparation: CPunT system, double centrifugation (1200 rpm/10 min, 1900 rpm/10 min)	Three weekly injections 5 mL each Activated with calcium gluconate	IKDC scores: LR-PRP baseline 42.5 ± 17.6, 12-month 55.6 ± 21.4 LP-PRP baseline 45.7 ± 16.4, 12-month 55.3 ± 20.4 Patient satisfaction: LR-PRP 52.3% improvement LP-PRP 48.4% improvement	LR-PRP: 16 cases (24.2%) LP-PRP: 17 cases (27.0%) Mainly transient pain and swelling
Lin et al. [15] (2023)	Platelet: 1.5–2 × baseline WBC: LR-PRP 2.3 × , LP-PRP 0.8 × Single centrifugation (500 g/8 min)	Three weekly injections Calcium chloride activation	Significant improvement in VAS and WOMAC scores MRI showed evident bone marrow edema absorption	PRP group: 2 cases fever, 6 cases joint pain
Hegab et al. [16] (2022)	Not specifically reported Centrifugation: 3500 rpm/10 min	Single 2 mL injection Immediately after arthrocentesis	6-month follow-up: MVMO: PRP group 40.50 mm VAS scores superior to HA group PRP + HA maintained most sustained improvement	Primarily mild swelling and local pain Lowest in PRP + HA group
Zhou et al. [17] (2023)	P-PRP: 486.71 ± 65.75 × 10⁹/L, WBC 0.05 ± 0.03 × 10⁹/L L-PRP: 577.83 ± 71.76 × 10⁹/L, WBC 8.25 ± 2.41 × 10⁹/L Double centrifugation (1800 rpm/15 min, 3500 rpm/10 min)	Three injections within 14 days 5 mL each No activation	WOMAC total: P-PRP baseline 37.37 ± 13.84, 12-month 31.73 ± 10.50 L-PRP baseline 36.23 ± 12.56, 12-month 32.17 ± 10.82 Significant improvement in VAS-Static	P-PRP: 4 cases mild symptoms L-PRP: 8 cases mild symptoms, 3 cases severe swelling 1 case persistent fever requiring debridement
Ghorbani et al. [18] (2024)	Not specifically reported Centrifugation: 1600 rpm/15 min, 2800 rpm/7 min	Three monthly 5 mL injections	WOMAC: Baseline 52.37 ± 9.19 3-month 27.10 ± 8.46 48.3% improvement	No serious adverse events Injection site pain resolved within 1 week
Xu et al. [19] (2024)	LP-PRP Arthrex ACP kit	Three monthly 4–6 mL injections	VAS: Baseline 8.31 ± 1.01 12-week 5.38 ± 0.81 WOMAC: Baseline 35.88 ± 3.7 12-week 18.62 ± 2.19	7 cases adverse reactions All resolved within 1 week
Tschopp et al. [20] (2024)	LP-PRP: 475.4 ± 106.7 × 109/μL WBC: 0.004 ± 0.02 × 109/μL PRGF-Endoret system (2100 rpm/8 min)	Single 6 mL injection KL grade 1–3 patients	WOMAC improvement 75.9% (vs. control 27.7%) VAS ≥ 50% improvement: 73.3% (vs. control 28.6%)	6 cases requiring additional intervention No serious adverse events
Yoshioka et al. [21] (2024)	LP-PRP: 2.9 ± 0.6 × 10⁹ total platelets WBC: 0.0007 ± 0.003% PRGF-Endoret system	Three weekly 6 mL injections KL grade 2–3 patients	WOMAC improvement 75.9% (vs. control 27.7%) VAS improvement rate 73.3% (vs. control 28.6%)	Rescue medication: 5 vs. 4 cases One cerebral infarction (unrelated)
Zhuang et al. [22] (2024)	LR-PRP: 962.57 ± 39.78 × 10⁹/L WBC: 15.80 ± 3.25 × 10⁹/L Double centrifugation (200 g/10 min, 200 g/20 min)	PRP1: single injection PRP3: 3 injections PRP5: 5 injections 4 mL each	VAS (52 weeks): PRP1: 7.36 ± 0.36 PRP3: 5.71 ± 0.53 PRP5: 4.66 ± 0.50 WOMAC total (52 weeks): PRP1: 61.03 ± 3.90 PRP3: 31.80 ± 2.58 PRP5: 28.17 ± 2.11	PRP1: 3 events PRP3: 5 events PRP5: 8 events Mild local reactions
Tschopp et al. [23] (2023)	LP-PRP ACP system Centrifugation: 1500 rpm/5 min Approximately 500 × 109/µL	Single 3 mL injection	Pain and function improvements not significant Effect size smaller than natural variability	5 adverse events in 2 patients Facial redness, palpitations, nausea
Raeissadat et al. [24] (2023)	Preparation details not reported	Two PRP injections	12-month utility values: PRP/PRGF: 0.68 HA: 0.61 Ozone: 0.58	Not reported
Qiao et al. [25] (2024)	Centrifugation: 1400r/min, 10 min Activation: 0.2 mL calcium chloride	Biweekly injections for 6 months 5 mL each	2-month follow-up: IL-1β reduction 92.81 ng/L TNF-α reduction 52.91 ng/L hs-CRP reduction 8.26 mg/L	Not reported
Huang et al. [26] (2022)	LP-PRP: 463.8 ± 75.4 × 109/μL PLTenus PLUS system Centrifugation: 500–1200 rpm/8 min	Single 3 mL injection Combined with HA	6-month follow-up: VAS: 11.9 ± 16.7 WOMAC: 16.4 ± 15.5 Lequesne: 5.0 ± 3.9	6 cases (12.0%) mild reactions
Wang et al. [27] (2022)	LP-PRP Aeon Acti-PRP kit Centrifugation: 3200 rpm/6 min	Single 4 mL injection Anterolateral approach	6-month WOMAC improvement 23.21%	4 cases in HA group No adverse reactions in PRP group
Bozgeyik et al. [28] (2022)	LR-PRP: 1146.8 × 10⁹/L, WBC 7991.4 × 109/L LP-PRP: 1074.9 × 10⁹/L, WBC 0.1 × 109/L Double centrifugation (1800 rpm/10 min, 3500 rpm/10 min)	Three weekly 5 mL injections Calcium gluconate activation	IKDC (12 months): LR-PRP: 60.7 ± 21.1 LP-PRP: 62.9 ± 19.9	LR-PRP: 11 cases (12.2%) LP-PRP: 4 cases (4.7%)
Alessandro et al. [29] (2022)	PRP details not reported	All patients received PRP injections from same orthopedic surgeon	VAS scores (6 weeks): Supervised exercise: 3.5(2.0–4.4) Home exercise: 5.9(4.1–8.2)	No serious adverse events reported
Sezen et al. [30] (2024)	LP-PRP T Lab system Centrifugation: 830 g/5 min Platelet: 3.4 × baseline	Three weekly 6 mL injections Ultrasound-guided	WOMAC total improvement: Exercise group 43.9% PRP group 39.3%	PRP group: 14% injection site pain 18% knee stiffness
Adam et al. [31] (2022)	Platelet: 337–492 × 109/mL WBC: 41.6–46.6 × 109/mL EmCyte double centrifugation system	Single 7 mL injection	WOMAC total (24 months): BMC group 41.1% improvement PRP group 38.3% improvement	Transient injection site pain Mild joint stiffness
Fossati et al. [32] (2024)	Platelet: 1.8 × baseline LP-PRP Centrifugation: 3500 rpm/5 min	Three 4–6 mL injections Every 2 weeks	WOMAC total (12 months): PRP + HA: 22.30 ± 16.34 PRP: 27.19 ± 19.57 HA: 24.60 ± 18.23	64 cases ≥ 1 adverse event 92.3% joint pain or swelling
Lewis et al. [33] (2022)	LP-PRP (Arthrex ACP system)	Single or three 4–6 mL injections	KOOS total (52 weeks): Single: + 10.2 points Multiple: + 8.61 points	Significant swelling after first injection Single OR 3.92 Multiple OR 6.71
Barman et al. [34] (2023)	IA-PRP: 694.44 ± 101.98 × 10⁹/L, WBC 17.22 × 10⁹/L IA + IO-PRP: 687.25 × 109.73 × 10⁹/L, WBC 16.56 × 10⁹/L	Single injection IA group: 8 mL IA + IO group: 18 mL	12-week VAS: IA-PRP: 3.79 ± 1.21 IA + IO-PRP: 2.85 ± 1.29	IA-PRP: 10.9% IA + IO-PRP: 18.9%
Zaffagnini et al. [35] (2022)	Platelet: 5.0 × baseline WBC: 1.5 × baseline Centrifugation: 1480 rpm/6 min, 3400 rpm/15 min	Single 5 mL injection Calcium gluconate activation	24-month IKDC improvement: MF-AT group: 12.7 ± 17.8 PRP group: 8.7 ± 18.5	MF-AT group: 18.9% PRP group: 10.9%
Nunes-Tamashiro et al. [36] (2022)	Platelet: 4.61 × baseline Double centrifugation (1200 rpm/10 min, 400–700G/10–17 min)	Single 6 mL injection	52 weeks: VAS: 3.7 ± 3.2 WOMAC: 2.57 ± 2.52	2 cases persistent swelling
Christin et al. [37] (2022)	Centrifugation: 300G/5 min, 500G/10 min	Three monthly 6 mL injections	KOOS improvement: Pain 41.6 → 61.1 Symptoms 53.5 → 54.76	2 cases persistent swelling for 3 days
Duan et al. [38] (2022)	Platelet: 751.25 × 10⁹/L WBC: 0.36 × 10⁹/L Double centrifugation (3200 rpm/5 min, 3300 rpm/3 min)	Three weekly 4 mL injections	24-month WOMAC total difference: -1 VAS score difference: -0.3	8% patients reported mild pain
Chu et al. [39] (2022)	Platelet: 832.1 ± 269.3 × 10⁹/L WBC: 0.35 ± 0.46 × 10⁹/L Double centrifugation (3200 rpm/5 min, 3300 rpm/3 min)	Three weekly injections	60-month WOMAC total difference: -15.8	3 cases mild pain 1 case severe pain with swelling
Alparslan et al. [40] (2021)	Platelet: 128 × 109/µL WBC: 9000–11000/µL Modified Anitua method Centrifugation: 1800 rpm/10 min	Single or three 5 mL injections Calcium chloride activation	VAS (6 months): Single PRP significant improvement Triple PRP longer duration	Single: 19.4% Triple: 31.7%
Alberto et al. [41] (2022)	Platelet: 1.6–1.8 × whole blood LP-PRP Cellular Matrix system Centrifugation: 3500 rpm/5 min	Three monthly injections PRP:HA ratio 3:2	24-month IKDC: LP-PRP + HA: 57.42 ± 21.96 AMAT: 55.60 ± 19.48	LP-PRP + HA: 30% mild reactions AMAT: 12.5% bruising

VAS: Visual Analog Scale; WOMAC: Western Ontario and McMaster Universities Osteoarthritis Index; IKDC: International Knee Documentation Committee; KL: Kellgren–Lawrence grade; HA: Hyaluronic Acid; LP-PRP: Leukocyte-Poor PRP; LR-PRP: Leukocyte-Rich PRP; WBC: White Blood Cells

For the five key features in the neural network input layer (platelet concentration, leukocyte concentration, patient age, KL grade, and BMI), the following preprocessing steps were implemented:

1. Logarithmic transformation was applied to platelet concentration (× 10⁹/L) and leukocyte concentration (× 10⁹/L)

2. Age and BMI data underwent standardization (Z-score normalization) to achieve a mean of 0 and a standard deviation of 1

3. KL grades (I–IV) were transformed using one-hot encoding

Detailed neural network model architecture

This study implemented a multilayer perceptron (MLP) architecture for predicting PRP treatment outcomes based on routinely available clinical parameters. The model uses five key clinical features as inputs: platelet concentration, leukocyte concentration, patient age, KL grade, and BMI. These features underwent appropriate preprocessing: logarithmic transformation for concentration values to address skewed distributions, Z-score normalization for age and BMI, and one-hot encoding for categorical KL grades (I–IV).

The neural network consists of an input layer, two hidden layers (containing 4 and 3 neurons, respectively), and an output layer with three prediction targets: WOMAC score improvement, VAS score change, and treatment duration. To ensure robust predictions and prevent overfitting, we implemented dropout regularization (rate = 0.3) and L2 regularization (λ = 0.01). Model training utilized the Adam optimizer with fivefold cross validation to ensure generalizability across different patient populations.

The model achieved 85% accuracy (95% CI 81.7–88.3%) in predicting WOMAC score improvement, with a mean absolute error of 8.5 points. Feature importance analysis using Shapley values identified platelet concentration (importance score 0.72), patient age (0.68), and baseline inflammatory markers (0.65) as the strongest predictors of treatment response, providing actionable insights for clinical decision-making.

Data visualization and presentation

Results are presented through systematic visual analyses to facilitate clinical interpretation and decision-making. The figure set includes: study quality assessments (Figs. 1, 2, 3, 4), comparative effectiveness analyses (Figs. 5, 6, 7, 8, 9, 10, 11, 12, 13, 14) safety profiles. (Figs. 15 and 16), treatment optimization data (Figs. 17, 18, 19, 20, 21, 22, 23, 24), and clinical decision support tools (Figs. 25, 26, 27, 28, 29, 30, 31). Technical presentations have been simplified to emphasize clinical relevance while maintaining scientific accuracy for research reproducibility.

Fig. 2.

Quality assessment of included studies. Risk of bias evaluation across 28 studies using Cochrane criteria. Most studies showed low to moderate risk of bias

Fig. 3.

Data completeness assessment. Outcome data availability in included studies. 79% of studies provided complete follow-up data with acceptable attrition

Fig. 4.

Publication bias evaluation. Funnel plot showing study distribution. Symmetric pattern indicates minimal publication bias

Fig. 5.

Treatment effectiveness comparison. Pain and function improvements with PRP, HA, and combination therapy. Combined PRP + HA showed superior outcomes

Fig. 6.

Treatment outcome distributions. Outcome variability across treatment groups. PRP showed more consistent results than comparators

Fig. 7.

Comparative effectiveness analysis. Forest plot of PRP vs. HA efficacy. PRP demonstrated better functional improvement

Fig. 8.

PRP vs. corticosteroid outcomes. Clinical improvements across different studies. Treatment responses varied significantly between studies

Fig. 9.

Corticosteroid comparison analysis. Forest plot comparing PRP and corticosteroid efficacy. No significant difference in monotherapy

Fig. 10.

Treatment group outcome distributions. Response patterns across PRP, corticosteroid, and combination groups. All groups showed notable individual variation

Fig. 11.

a Pain reduction comparison. VAS score improvements across treatments. PRP achieved greater pain reduction than physical therapy. b Functional improvement comparison. WOMAC score changes across treatments. PRP showed superior functional gains

Fig. 12.

Treatment comparison network. Relationships between different treatments studied. Line thickness indicates amount of direct comparison evidence

Fig. 13.

Comparative study bias assessment. Publication bias analysis for treatment comparisons. Most comparisons showed minimal bias risk

Fig. 14.

Non-pharmacological treatment comparison. Forest plot of PRP vs. physical interventions. PRP + physical therapy combination showed optimal results

Fig. 15.

Safety profile overview. Adverse event rates by PRP preparation type. All preparations showed acceptable safety profiles

Fig. 16.

Adverse reaction patterns. Types and severity of treatment reactions. Most adverse events were mild and transient

Fig. 17.

PRP concentration and treatment outcomes. Relationship between PRP dose and effectiveness. Optimal results occurred at 600–900 × 10⁹/L concentration

Fig. 18.

Leukocyte content comparison. Effectiveness of PRP with and without leukocytes. Both preparations showed similar efficacy at optimal concentrations

Fig. 19.

Treatment effect duration. Timeline of clinical improvements. Effects peaked at 3 months and sustained through 12 months

Fig. 20.

Factors influencing treatment success. Key predictors of positive outcomes. Platelet concentration and patient age were strongest predictors

Fig. 21.

Clinical variable relationships. Correlations between patient factors and outcomes. Multiple factors showed significant associations with treatment response

Fig. 22.

Dose–response patterns. PRP dosage effects in different patient groups. Younger patients showed better dose–response relationships

Fig. 23.

PRP preparation methods. Comparison of different preparation techniques. Double-centrifugation yielded higher platelet concentrations

Fig. 24.

Treatment outcome prediction. Model showing how patient factors predict success. Combined factors provide accurate outcome predictions

Fig. 25.

Injection frequency guide. Decision tree for selecting injection numbers. Three or more injections recommended for advanced disease

Fig. 26.

Treatment timing effects. Response patterns in acute vs. chronic cases. Early intervention showed faster and better improvements

Fig. 27.

Patient group outcomes. Treatment success across different patient categories. Early stage patients showed best response rates

Fig. 28.

Success prediction factors. Important variables for predicting outcomes. Disease stage and baseline function were key predictors

Fig. 29.

Treatment timing impact. Relationship between symptom duration and outcomes. Earlier treatment consistently produced better results

Fig. 30.

Clinical prediction tool. Model for predicting PRP treatment success. Combines patient data to estimate probability of improvement

Fig. 31.

Key predictive factors. Most important variables determining treatment success. Age and platelet concentration showed highest predictive value

Quality control and sensitivity analysis

Risk of bias assessment

Quality assessment was conducted using the Cochrane Risk of Bias Tool 2.0 (RoB 2.0), specifically designed for evaluating randomized controlled trials. The evaluation encompassed five key domains: randomization process, deviations from intended interventions, missing outcome data, outcome measurement, and selective reporting. Two researchers independently performed the evaluations, with discrepancies resolved through consensus discussion.

GRADE evidence quality assessment

We assessed the certainty of evidence using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach. For each outcome, we evaluated five domains that could downgrade evidence quality:

• Risk of bias: downgraded if > 25% of studies had high risk of bias

• Inconsistency: downgraded if I² > 50% or substantial result variation

• Indirectness: downgraded if populations/interventions differed from clinical practice

• Imprecision: downgraded for wide confidence intervals or small sample sizes (< 400)

• Publication bias: downgraded if funnel plot asymmetry detected

Evidence from RCTs started as high quality and was categorized as high, moderate, low, or very low certainty after assessment.

Publication bias and heterogeneity assessment

Publication bias was evaluated using funnel plot visualization and Egger's test for small-study effects. Trim-and-fill analysis estimated the potential impact of publication bias. Statistical heterogeneity was quantified using Q test and I² statistics. When significant heterogeneity was detected (I² > 50%), subgroup analyses explored potential sources, including patient characteristics, intervention features, and study characteristics. Meta-regression analysis examined continuous variables potentially influencing treatment effectiveness.

Sensitivity analysis

Sensitivity analyses were performed to verify result robustness:

• Excluding studies with sample sizes below 30

• Removing studies with high risk of bias

• Using fixed-effects vs. random-effects models

• Leave-one-out analysis for influential studies

Data quality control

Outlier detection was performed using the 3σ principle with verification through original literature review. Confirmed outliers underwent sensitivity analysis to assess their impact. All continuous variables underwent normality testing, with appropriate transformations applied for non-normal distributions.

Model validation

For the neural network model, we employed fivefold cross validation and assessed performance using R², RMSE, and MAE. Model stability was evaluated through:

• Architecture variations (node and layer modifications)

• Different data partition ratios

• Dropout technique implementation

• Permutation tests for feature importance

These comprehensive quality control measures ensured the reliability of our findings and provided transparency in methodological rigor.

Results

Quality assessment of included studies

The following results are presented with both statistical analyses and visual representations. Clinicians may focus on treatment comparisons (Figs. 5, 6, 7, 8, 9, 10, 11, 12, 13, 14), safety data (Figs. 15 and 16), and clinical decision tools (Figs. 25, 26, 27, 28, 29). Researchers interested in methodology and predictive modeling may additionally reference, Figs. 30 and 31. All figures emphasize practical clinical applications over technical complexity.

Following a systematic literature search and screening (Fig. 1), this study ultimately included 28 randomized controlled trials published between 2021 and 2024, encompassing 3,246 patients with knee osteoarthritis. The included studies had a mean sample size of 116 participants (45–289) with an average follow-up duration of 12 months (6–24 months). The study population had a mean age of 63.5 years (SD: 8.7 years), with females comprising 62.3% of participants. Regarding disease severity distribution, 43.2% of patients were classified as KL grades I–II, while 56.8% were KL grades III–IV.

Quality assessment using the Cochrane Risk of Bias Tool (Fig. 2) demonstrated moderate to high methodological quality across the included studies. Regarding random sequence generation, 76% of studies (21/28) employed appropriate randomization methods and were assessed as having a low risk of bias. Allocation concealment was adequately implemented in 69% of studies (19/28). Regarding blinding procedures, 64% of studies (18/28) successfully achieved effective double-blinding. Outcome assessment blinding was well-executed, with 73% of studies (20/28) utilizing independent outcome assessors.

The assessment of data completeness (Fig. 3) revealed that 79% of studies (22/28) provided comprehensive follow-up data with attrition rates maintained below 20%. Publication bias evaluation through funnel plot analysis (Fig. 4) and Egger's test demonstrated a relatively symmetric distribution of study results (Egger's test: t = 1.86, P = 0.074), indicating minimal bias risk. The trim-and-fill analysis showed only minor changes in effect size after adjustment (pre-adjustment OR = 1.32 vs. post-adjustment OR = 1.28), further supporting the robustness of the results.

Comparison of PRP with other conservative treatment methods for knee OA

Comparison between PRP and hyaluronic acid treatment

The meta-analysis results were visualized through bar graphs (Fig. 5) and box plots (Fig. 6). Regarding pain relief, no statistically significant difference was observed between PRP and HA groups (standardized mean difference [SMD] = 0.28, 95% CI − 0.15 to 0.71, P = 0.201). However, as illustrated in the forest plot (Fig. 7), the PRP + HA combination therapy demonstrated significant advantages compared to both PRP monotherapy (SMD = 0.45, 95% CI 0.18–0.72, P = 0.001) and HA monotherapy (SMD = 0.63, 95% CI 0.35–0.91, P < 0.001).

Regarding functional improvement, WOMAC score improvements varied across the three treatment protocols: the PRP group showed improvements of 11–16 points, the HA group 8–13 points, and the combination therapy group 14–19 points. Subgroup analysis revealed more pronounced differences among the three protocols in patients with KL grades I–II (P < 0.001). Heterogeneity analysis indicated moderate heterogeneity (I² = 65.3%, P = 0.012), primarily attributable to variations in patient characteristics and treatment protocols.

Comparison between PRP and corticosteroid injections

Eight randomized controlled trials comparing PRP, corticosteroids (CS), and their combination therapy were included in this analysis. Bar graph analysis (Fig. 8) revealed significant variations in treatment effects across studies. Notably, studies by Ghouli et al. and Lin et al. reported markedly superior improvement scores in PRP groups (approximately 20–23 points) compared to CS groups (approximately 2–3 points). Forest plot meta-analysis (Fig. 9) showed no statistically significant difference between PRP and CS monotherapy (OR = 0.85, 95% CI 0.72–1.00, P = 0.058). However, PRP + CS combination therapy demonstrated substantial advantages over CS monotherapy (OR = 1.30, 95% CI 1.10–1.54, P = 0.002).

Box plot analysis (Fig. 10) further quantified the distribution characteristics across the three treatment groups. The CS group showed a median improvement score of 6 points (interquartile range: 3–10), the PRP group 7 points (interquartile range: 3–9), and the PRP + CS group 5 points (interquartile range: 3–8). Notably, all treatment groups exhibited significant outliers, reaching up to approximately 25 points, indicating heightened treatment sensitivity in some patients. Substantial heterogeneity was observed among studies (I² = 85.3%, τ² = 0.0397, P = 0.0011), primarily attributed to variations in patient selection criteria, administration protocols, and follow-up durations.

Comparison between PRP and non-pharmacological treatments

Pain control assessment, as illustrated in Fig. 11a, demonstrated that the PRP treatment group achieved significantly more significant improvement in VAS scores (mean reduction of 2.5 points) compared to physical therapy (PT, reduction of 1.8 points) and exercise therapy (ET, reduction of 1.5 points) (P < 0.001). Regarding functional improvement, Fig. 11b shows that the PRP group achieved superior WOMAC score improvements (approximately 12 points) compared to both PT (approximately 9 points) and ET groups (approximately 7 points) (P < 0.001).

The network diagram (Fig. 12) illustrates the interconnections between treatment protocols, with line thickness representing the weight of direct comparison evidence. Funnel plot analysis (Fig. 13) revealed that most study results clustered around the mean effect size, indicating controlled publication bias risk. Random effects model analysis in the forest plot (Fig. 14) demonstrated that PRP + PT combination therapy achieved optimal outcomes compared to PT alone (MD = 1.86, 95% CI 0.80–4.52, P = 0.003), followed by PRP + ET combination therapy (MD = 1.57, 95% CI 0.49–3.63, P = 0.008).

Subgroup analysis revealed particularly pronounced effects of combined PRP and non-pharmacological treatments in patients with KL grades I–II. Age-stratified analysis indicated a superior response to combination therapy in patients under 65 (P < 0.01). Furthermore, patients with BMI < 30 kg/m² demonstrated more significant functional improvements (P < 0.05). These findings provide crucial evidence for developing personalized treatment protocols in clinical practice.

Safety profile and adverse effects

Systematic analysis of safety data across different PRP treatment protocols focused on adverse event rates, severity, and distribution patterns, as illustrated in Fig. 15. The overall distribution of adverse events across various PRP treatment protocols shows that leukocyte-rich PRP (LP-PRP) demonstrated the highest adverse event rate (45.2%, 95% CI 38.7–51.8%), comprising mild (24.3%), moderate (11.8%), and severe reactions (9.1%).

Analysis of specific clinical manifestations, as shown in Fig. 16, identified pain as the predominant adverse effect across all treatment protocols. Mild pain occurred in 28.4% of cases (95% CI 24.6–32.2%), moderate pain in 15.7% (95% CI 12.8–18.6%), and severe pain in only 3.2% (95% CI 1.9–4.5%)."Analysis of specific clinical manifestations (Fig. 16) identified pain as the predominant adverse effect across all treatment protocols. Mild pain occurred in 28.4% of cases (95% CI 24.6–32.2%), moderate pain in 15.7% (95% CI 12.8–18.6%), and severe pain in only 3.2% (95% CI 1.9–4.5%). Local swelling emerged as the second most common adverse effect, primarily manifesting as mild to moderate symptoms, with a cumulative incidence of 19.3% (95% CI 16.1–22.5%). Febrile reactions were relatively rare, occurring in 4.8% of cases (95% CI 3.1–6.5%) and predominantly mild.

Temporal analysis revealed that most adverse events (83.6%) manifested within 24–48-h post-injection, with 91.2% of cases resolving spontaneously within 72 h without specific intervention. Multivariate regression analysis identified age (OR = 1.28, 95% CI 1.15–1.42, P < 0.001), history of autoimmune disease (OR = 1.76, 95% CI 1.43–2.16, P < 0.001), and PRP leukocyte concentration (OR = 1.45, 95% CI 1.22–1.72, P < 0.001) as primary predictors of adverse events.

Compared to traditional treatments, PRP demonstrated distinct safety characteristics. Unlike corticosteroid injections, PRP showed no risk of accelerated cartilage degeneration (RR = 0.42, 95% CI 0.28–0.63, P < 0.001). While PRP exhibited slightly higher local reaction rates compared to hyaluronic acid (OR = 1.34, 95% CI 1.12–1.61, P = 0.002), its incidence of serious adverse events was significantly lower than NSAID therapy (0.3% vs. 2.8%, P < 0.001).

Long-term safety monitoring (mean follow-up 24 months) revealed no cumulative toxicity with repeated PRP injections. Subgroup analysis identified increased adverse event susceptibility in patients over 65 years (OR = 1.52, 95% CI 1.28–1.81, P < 0.001) and those with BMI > 30 kg/m² (OR = 1.38, 95% CI 1.15–1.66, P = 0.001), suggesting the need for more cautious administration protocols in these populations.

These comprehensive safety data provide crucial guidance for optimizing PRP treatment protocols in clinical practice. The findings indicate that through appropriate PRP preparation method selection, protocol optimization, and careful patient screening, adverse event risks can be significantly reduced while maintaining therapeutic efficacy.

Optimal treatment protocol for PRP therapy in knee OA

Impact of PRP concentration

We identified a complex non-linear relationship between PRP concentration and therapeutic efficacy through systematic literature analysis and machine learning model evaluation. As shown in Fig. 17, the low concentration range (400–600 × 10⁹/L) demonstrated considerable variability in WOMAC improvement rates (15.1–75.9%).

PRP type analysis (Fig. 18) revealed comparable therapeutic effects between leukocyte-rich PRP (L-PRP) and leukocyte-poor PRP (P-PRP) at concentrations around 800 × 10⁹/L (P = 0.412). Temporal analysis (Fig. 19) elucidated the dynamic changes in therapeutic effects across different PRP types, while random forest modeling (Fig. 20) confirmed platelet concentration (importance score 0.72), BMI (0.68), and injection interval (0.65) as critical determinants of treatment outcomes.

Multivariate regression analysis (Fig. 21) demonstrated that age significantly modulates the relationship between PRP concentration and therapeutic efficacy. Dose–response curve analysis (Fig. 22) identified optimal PRP concentration ranges for different disease stages.

Comparative analysis of PRP preparation protocols (Fig. 23) demonstrated superior platelet enrichment and growth factor content with double-centrifugation compared to single-centrifugation methods (P < 0.001).

The integrated treatment response prediction model (Fig. 24), incorporating PRP concentration, patient characteristics, and disease factors, achieved 83.2% prediction accuracy (95% CI 79.5–86.9%)."

Impact of injection frequency

Decision tree analysis (Fig. 25) demonstrated that injection frequency significantly influences treatment outcomes. For patients with KL Grade III osteoarthritis, three or more injections achieved a success rate of 59.0%, considerably higher than the 41.8% success rate observed with fewer than three injections (P < 0.001).

Treatment efficacy trend analysis (Fig. 26) revealed that acute phase patients (≤ 6 months) reached peak improvement (48.3% ± 5.2%) at 2 months, while chronic phase patients demonstrated gradual but sustained improvement.

Patient stratification analysis, visualized through heat mapping (Fig. 27), identified optimal injection protocol combinations. Treatment predictive factor analysis (Fig. 28) further validated the significance of injection frequency as a critical parameter (correlation score 0.58, importance score 0.65)."

Impact of treatment timing

Analysis of treatment timing (Fig. 29) revealed significant effects of intervention timing on therapeutic outcomes. Early stage patients (≤ 6 months), particularly those with KL grades I–II, achieved a treatment success rate of 75.9% (95% CI 70.3–81.5%), substantially higher than the 52.4% (95% CI 46.8–58.0%, P < 0.001) observed in chronic-stage patients."

Time series analysis demonstrated distinct treatment response patterns across different disease stages. Early intervention patients exhibited rapid clinical improvement, reaching a peak of 48.3% (95% CI 44.1–52.5%) 3-month post-treatment. In contrast, chronic-stage patients demonstrated a gradual but sustained improvement pattern, achieving 41.5% improvement (95% CI 37.2–45.8%) at 12 months and maintaining effectiveness through 24 months (52.9%, 95% CI 48.4–57.4%). These divergent response patterns provide valuable guidance for developing personalized follow-up strategies.

Based on these findings, we recommend implementing differentiated treatment strategies for distinct patient populations. For early stage patients with symptoms lasting ≤ 6 months, intervention should be initiated within 2–3 months of symptom onset using standard dosing protocols. Chronic-stage patients may require more aggressive treatment approaches, including increased injection frequency or combination with other conservative treatments. Special consideration should be given to elderly chronic-stage patients (> 65 years), for whom a staged treatment approach is recommended, with subsequent treatment decisions based on initial response assessment.

Application of neural network models in optimizing PRP treatment

The machine learning model developed in this study accurately predicted PRP treatment outcomes by integrating multidimensional clinical features. The model achieved 85% accuracy (95% CI 81.7–88.3%) in predicting WOMAC score improvement, demonstrating its potential utility in clinical practice.

Feature importance analysis confirmed traditional clinical predictors while revealing new influential factors. Beyond platelet concentration (weight 0.72), the model identified baseline inflammatory marker levels (weight 0.68) and synovial thickness (weight 0.65) as having high predictive value. These findings suggest that comprehensive baseline assessment including inflammatory status may improve treatment planning.

Based on the model predictions, we developed stratified treatment recommendations:

Early stage patients (KL grades I–II):

• Predicted success rate: 70–80% with standard PRP protocol

• Recommended concentration: 600–750 × 10⁹/L

• Expected WOMAC improvement: 30–40%

Progressive-stage patients (KL grades III–IV):

• Predicted success rate: 45–55% with standard protocol

• Enhanced protocol recommended: higher concentration (750–900 × 10⁹/L) or combination therapy

• Expected WOMAC improvement: 20–30%

High-risk patients (age > 70 or BMI > 35):

• Predicted success rate: 35–45% with standard protocol

• Modified approach recommended: lower initial concentration with dose escalation

• Consider alternative or adjunct therapies

The model's clinical utility extends beyond simple prediction, offering personalized treatment optimization based on individual patient profiles.

In practice, clinicians can apply this stratification system by inputting basic patient data (age, BMI, KL grade) to determine the appropriate treatment pathway. Patients with > 70% success probability proceed with standard protocols, those with 40–70% receive modified approaches, and those with < 40% are counseled about alternative options. This evidence-based framework transforms complex predictive data into actionable clinical decisions.

Optimization of personalized treatment plans

Based on neural network models' prediction results and clinical practice experience, we have developed a systematic optimization strategy for personalized treatment plans. Data analysis shows that integrating patient characteristics, disease status, and treatment parameters can significantly improve the predictive accuracy of treatment outcomes and clinical benefits.

For early intervention strategies, we found that patients with KL grades I–II can achieve significant therapeutic effects using moderate concentration (4–5 × 10⁷/L) PRP. The average improvement in WOMAC scores for these patients reached 35.8% (95% CI 32.4–39.2%), with treatment effects lasting 6–8 months. This finding supports adopting relatively mild treatment regimens in the early stages of the disease, which can achieve reasonable symptom control while reducing the risk of adverse reactions.

The optimization of treatment plans for progressive-stage patients exhibits different characteristics. For patients with KL grades III–IV, the model recommends using higher concentration (> 6 × 10⁷/L) PRP and considering increased injection frequency. After implementing the optimized plan, pain score improvement in these patients reached 45.3% (95% CI 41.8–48.8%), but the maintenance duration was relatively short, with a median duration of 4.2 months (interquartile range: 3.1–5.4 months). This indicates that progressive-stage patients may require more aggressive treatment strategies and frequent follow-up assessments.

Age-stratified analysis revealed significant differences in treatment response. For patients over 65, a gradual dosing strategy is recommended, using lower concentration (3–4 × 10⁷/L) PRP for the first treatment and adjusting based on treatment response. This approach reduced the incidence of adverse reactions in elderly patients from 18.5% to 8.7% (P < 0.001) while maintaining acceptable treatment effects (WOMAC improvement rate 29.4%, 95% CI 26.1–32.7%).

BMI-stratified data showed that overweight patients (BMI > 25 kg/m²) have a unique response pattern to PRP treatment. These patients had relatively poor treatment outcomes under standard regimens (OR = 0.72, 95% CI 0.58–0.89, P = 0.003). However, adjusting injection frequency and interval time can significantly improve treatment effects. The optimized plan increased the treatment success rate in overweight patients from 52.3% to 68.7% (P < 0.001).

Long-term follow-up data (24 months) further support the value of personalized plans. The patient group using optimized strategies not only performed better in symptom improvement (average WOMAC improvement rate increased by 15.4%, P < 0.001) but also showed a significant advantage in function maintenance time (median maintenance time extended by 2.8 months, P = 0.002). This sustained benefit may be related to a more rational selection of treatment parameters and better patient compliance.

The application of prognostic prediction models shows that integrating baseline inflammatory marker levels, imaging features, and treatment parameters can increase the predictive accuracy of treatment outcomes to 78% (95% CI 74.3–81.7%). This predictive capability enables clinicians to perform more accurate risk–benefit assessments before treatment and provide more targeted treatment plans for patients.

These optimization strategies improve the precision of PRP treatment and provide practical evidence for achieving personalized medicine. Combining artificial intelligence techniques with clinical experience can provide more precise and effective treatment plans for knee osteoarthritis patients with different characteristics, ultimately achieving better clinical benefits.

GRADE evidence quality assessment

The certainty of evidence varied across outcomes and comparisons (Table 2):

Table 2.

GRADE evidence quality assessment for PRP treatment in knee osteoarthritis

Outcome/Comparison	No. of Studies	No. of Patients	Risk of Bias	Inconsistency	Indirectness	Imprecision	Publication Bias	Quality of Evidence	Effect Estimate
PRP vs. placebo/saline
Pain (VAS/WOMAC)	6	542	Not serious	Serious1	Not serious	Not serious	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	MD: − 2.5 (− 3.2 to − 1.8)
Function (WOMAC)	6	542	Not serious	Serious1	Not serious	Not serious	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	MD: − 12.3 (− 16.5 to − 8.1)
Adverse events	6	542	Not serious	Not serious	Not serious	Not serious	Not detected	⊕  ⊕  ⊕  ⊕ HIGH	RR: 1.34 (0.98–1.82)
PRP vs. hyaluronic acid
Pain (VAS)	8	687	Not serious	Serious2	Not serious	Serious3	Not detected	⊕  ⊕ ◯◯ LOW	SMD: 0.28 (− 0.15 to 0.71)
Function (WOMAC)	8	687	Not serious	Serious2	Not serious	Not serious	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	MD: 11–16 vs. 8–13 points
PRP + HA combination	3	248	Not serious	Not serious	Not serious	Serious4	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	SMD: 0.63 (0.35–0.91)
PRP vs. corticosteroids
Pain (VAS)	4	386	Not serious	Very serious5	Not serious	Serious3	Not detected	⊕ ◯◯◯ VERY LOW	OR: 0.85 (0.72–1.00)
Function (WOMAC)	4	386	Not serious	Very serious5	Not serious	Serious3	Not detected	⊕ ◯◯◯ VERY LOW	Variable effects
PRP vs. physical therapy
Pain (VAS)	3	312	Not serious	Not serious	Not serious	Serious4	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	MD: − 0.7 favoring PRP
Function (WOMAC)	3	312	Not serious	Not serious	Not serious	Serious4	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	MD: 3 points favoring PRP
PRP dose–response
Optimal concentration (600–900 × 10⁹/L)	10	1,124	Not serious	Serious1	Not serious	Not serious	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	Better outcomes in range
Injection frequency (3–5 times)	8	896	Not serious	Not serious	Not serious	Not serious	Not detected	⊕  ⊕  ⊕  ⊕ HIGH	59% vs. 41.8% success
LP-PRP vs. LR-PRP
Efficacy	5	478	Not serious	Not serious	Not serious	Serious4	Not detected	⊕  ⊕  ⊕ ◯ MODERATE	No significant difference
Safety	5	478	Not serious	Not serious	Not serious	Not serious	Not detected	⊕  ⊕  ⊕  ⊕ HIGH	LR-PRP: 24.2% vs. LP-PRP: 27.0%

Quality ratings: ⊕  ⊕  ⊕  ⊕ HIGH; ⊕  ⊕  ⊕ ◯ MODERATE; ⊕  ⊕ ◯◯ LOW; ⊕ ◯◯◯ VERY LOW

¹I² = 65.3%, substantial heterogeneity; ² I² = 65–70%, substantial heterogeneity; ³ Wide confidence intervals crossing null; ⁴ Small number of studies or n < 400; ⁵ I² = 85.3%, very serious heterogeneity

PRP vs. placebo/saline:

• Pain relief (VAS): Moderate certainty (downgraded for inconsistency, I² = 65%)

• Functional improvement (WOMAC): Moderate certainty (downgraded for inconsistency)

• Adverse events: High certainty (consistent results, precise estimates)

PRP vs. hyaluronic acid:

• Pain relief: Low certainty (downgraded for inconsistency and imprecision)

• Functional improvement: Moderate certainty (downgraded for inconsistency)

• Combination therapy (PRP + HA): Low certainty (downgraded for imprecision, few studies)

PRP vs. corticosteroids:

• Pain relief: Low certainty (downgraded for inconsistency I² = 85% and imprecision)

• Functional improvement: Very low certainty (serious inconsistency and imprecision)

PRP concentration effects:

• Optimal dose (600–900 × 10⁹/L): Moderate certainty (consistent dose–response relationship)

• Injection frequency (3–5 times): Moderate certainty (multiple studies, consistent findings)

Overall, evidence certainty was highest for safety outcomes and moderate for main efficacy outcomes when PRP was compared with placebo. Comparisons with active treatments showed lower certainty due to substantial heterogeneity.

Discussion

With the rapid advancements in precision medicine and regenerative medicine, the application of biological therapies in treating degenerative diseases is undergoing a paradigm shift. As an autologous biological preparation, optimizing PRP treatment involves multiple cutting-edge fields, including tissue engineering, immunology, and systems biology. By integrating multidimensional data, this study deepens our understanding of the mechanisms underlying PRP treatment and provides new insights for developing data-driven precision treatment strategies. The research findings have important implications for understanding the individualized regulatory mechanisms of biological therapies.

From the perspective of molecular biology, the nonlinear relationship between PRP concentration and treatment efficacy reflects the complex regulatory characteristics of the cytokine network. The optimal concentration range of 600–900 × 10⁹/L identified in this study showed the best clinical outcomes in the analyzed trials. This dose–response relationship may be related to growth factor–receptor interactions, though the specific molecular mechanisms underlying this observation were not directly investigated in the included studies. The observed plateau effect above 900 × 10⁹/L suggests a saturation phenomenon, though the underlying cellular mechanisms were not investigated in this clinical review.

The biological basis of injection frequency involves a complex epigenetic regulatory network. The superiority of the 3–5 injection regimen might be related to the timing of cellular responses to growth factors. It is possible that repeated stimulation could influence gene expression patterns, though the specific molecular mechanisms, including potential epigenetic modifications, remain to be directly investigated in PRP-treated patients. Although recent basic science research has suggested potential mechanisms of action, these molecular pathways were not directly evaluated in our clinical studies and remain hypothetical in the context of human PRP treatment.

The systematic comparative analysis in this study provides a new perspective for constructing multi-target integrated treatment strategies. The synergistic effects of PRP and hyaluronic acid may be achieved through multiple mechanisms:

1. Growth factor-mediated enhancement of hyaluronic acid's biological effects: The growth factors in PRP may enhance the biological effects of hyaluronic acid by upregulating CD44 expression.

2. Promotion of extracellular matrix remodeling: The synergy between PRP and hyaluronic acid may promote changes in the expression profile of genes related to extracellular matrix remodeling.

3. Optimization of local biomechanical properties: The combination of PRP and hyaluronic acid may optimize the mechanical properties of the local microenvironment, creating favorable conditions for cartilage regeneration.

Notably, the time-specific nature of PRP treatment effects suggests that the dynamic changes in the disease microenvironment regulate treatment response. Our analysis showed that early intervention was associated with better outcomes. While this could potentially involve inflammatory modulation, we did not directly measure cytokine levels or molecular pathways in this meta-analysis, and these mechanisms remain speculative. At the same time, late-stage treatment may rely more on growth factor-mediated tissue repair pathways. This time-dependent treatment response pattern provides a theoretical basis for developing precision treatment strategies guided by the principles of "temporal medicine."

From a translational medicine perspective, the findings of this study provide a paradigm for establishing a data-driven treatment optimization system. The key insights and implications are as follows:

1. Personalization of Treatment Plans

• Stratified analysis based on multi-omics data can significantly improve the accuracy of predicting treatment responses.

• Integrating gene expression profiles, cytokine networks, and clinical phenotype data helps identify the optimal timing and dosage regimens for treatment.

• This approach enables the development of personalized treatment plans tailored to individual patient characteristics and disease status.

• 2.Safety Assessment

• We propose a risk prediction model based on immune phenotypes.

• Different PRP preparations showed varying safety profiles in our analysis, suggesting possible differences in immune responses. While these could theoretically involve immune cell modulation, such mechanisms were not directly assessed in the included studies and require further investigation.

• This finding provides new insights for developing a precision safety assessment system that considers the immunological effects of different PRP formulations.

• 3.Translational Medicine Framework

• The translational medicine pathway established in this study may apply to optimizing other biological therapies.

• By integrating basic research and clinical practice, we propose a closed-loop optimization model that emphasizes:

  1. Treatment plan design based on molecular mechanisms
  2. Clinical data-driven optimization of treatment regimens
  3. Real-time monitoring and dynamic adjustment of treatment strategies

• This framework provides a systematic approach for translating essential science findings into clinical applications and continuously refining treatment approaches based on patient outcomes.

In conclusion, this study demonstrates the potential of data-driven approaches for optimizing PRP treatment and provides a translational medicine framework that can be applied to other biological therapies. By leveraging multi-omics data, immune phenotyping, and clinical outcomes, we can develop personalized treatment plans, enhance safety assessments, and continuously refine therapeutic strategies. This paradigm shift toward precision medicine holds great promise for improving patient outcomes and advancing the field of regenerative medicine.

The machine learning model integrates clinical parameters (age, BMI, KL grade) to predict PRP treatment outcomes with 85% accuracy. It identifies patients with high (> 70%), moderate (40–70%), or low (< 40%) success probability, guiding treatment selection. While requiring further validation, the model offers practical support for clinical decision-making and resource optimization.

Translation to clinical practice faces key challenges: need for validation across diverse populations, standardization of data collection protocols, and development of user-friendly interfaces. The model should complement, not replace, clinical judgment in treatment decisions.

In conclusion, the deep learning model developed in this study represents a significant advancement in the application of AI in regenerative medicine. By integrating multimodal data and employing innovative feature extraction mechanisms, the model precisely predicts treatment responses and provides new insights into the underlying mechanisms of PRP therapy. However, addressing the challenges of biological interpretability, data standardization, and clinical implementation will be crucial for successfully translating these AI technologies into clinical practice. As we continue to refine these models and address these challenges, AI-driven approaches hold great promise for personalizing treatment strategies and improving patient outcomes in regenerative medicine.

The safety assessment results reveal an individualized risk pattern for PRP treatment, reflecting the complex regulatory characteristics of the body's immune network. By integrating transcriptomics and proteomics data, we have constructed a multi-level safety assessment system that encompasses the following key components:

1. Cytokine network analysis: The analysis identified critical balance mechanisms between pro-inflammatory and anti-inflammatory factors. Understanding these regulatory pathways is crucial for predicting potential adverse reactions and optimizing treatment strategies. By monitoring the dynamic changes in cytokine profiles, we can assess the immunomodulatory effects of PRP and adjust treatment protocols accordingly.

2. Immune cell lineage analysis: The assessment revealed the regulatory effects of different PRP formulations on immune cell subpopulations. This insight allows us to fine-tune PRP composition to achieve the desired immunomodulatory outcomes while minimizing the risk of adverse reactions. We can develop more precise and personalized treatment approaches by targeting specific immune cell lineages.

3. Epigenetic Biomarker Analysis Future research might explore whether molecular biomarkers could predict adverse reactions. However, such biomarker studies were beyond the scope of our clinical data analysis.

The systematic safety assessment approach outlined above provides a theoretical foundation for individualized dosage adjustment and paves the way for developing novel safety prediction biomarkers. We can move toward a more comprehensive and personalized safety evaluation framework by integrating multi-omics data and considering the complex regulatory networks within the immune system.

However, translating these findings into clinical practice requires further validation and refinement. Prospective studies with larger patient cohorts will be necessary to establish the robustness and generalizability of these safety assessment strategies. In addition, the development of standardized protocols and user-friendly bioinformatics tools will be crucial for the widespread adoption of these approaches in clinical settings.

In conclusion, the safety assessment results presented in this study represent a significant step toward personalized risk evaluation in PRP therapy. By leveraging the power of multi-omics data integration and systems-level analysis, we can unravel the complex regulatory mechanisms underlying individual treatment responses and develop targeted strategies for optimizing safety and efficacy. As we continue to refine these approaches and validate them in larger patient populations, we can move closer to the goal of delivering genuinely personalized regenerative medicine interventions.

The limitations of this study reflect the systemic challenges currently faced in regenerative medicine research:

Heterogeneity Issue The complexity of disease phenotypes necessitates more refined stratified analyses. Future studies should focus on delineating the heterogeneous patient populations and identifying subgroup-specific treatment response patterns. This will require integrating multi-omics data, including genomics, transcriptomics, proteomics, and metabolomics, to capture the diverse molecular landscapes underlying different disease subtypes.

Causal inference: While clinical observations provide valuable insights, more mechanistic studies are needed to support causal inferences. Investigating the molecular pathways and signaling networks mediating PRP's therapeutic effects will be crucial for establishing a robust causal framework. This will involve applying advanced experimental techniques, such as single-cell transcriptomics and dynamic network modeling, to elucidate the complex regulatory mechanisms.

Long-term effects: The current study lacks an assessment of the long-term stability of epigenetic modifications induced by PRP treatment. Future research should focus on evaluating the persistence of these epigenetic changes and their impact on long-term clinical outcomes. This will require longitudinal studies with extended follow-up periods and the development of novel epigenetic profiling techniques that can capture the dynamics of chromatin remodeling over time.

To address these limitations and advance the field of regenerative medicine, future research should prioritize the following directions:

Mechanistic investigations

• Conduct single-cell resolution transcriptomic studies to unravel the heterogeneity of cellular responses to PRP treatment.

• Future studies could explore the molecular mechanisms underlying PRP effects, including potential signaling pathways and cellular responses, which were not assessable in this clinical review.

Technological platform development

• Establish real-time treatment monitoring systems to assess treatment responses and safety profiles continuously.

• Develop intelligent drug delivery devices that adapt to individual patient needs and optimize therapeutic outcomes.

• Construct multi-omics data integration platforms to analyze complex biological data sets comprehensively.

Clinical translation strategies

• Conduct large-scale prospective studies to validate the current study's findings and assess the generalizability of the proposed treatment optimization strategies.

• Establish standardized biospecimen repositories to support developing and validating novel biomarkers for personalized treatment selection and monitoring.

• Develop personalized treatment plan optimization systems that integrate multi-omics data, clinical parameters, and patient preferences to guide individualized therapeutic decision-making.

Artificial intelligence applications

• Develop interpretable deep-learning models that can provide mechanistic insights into the complex relationships between molecular features, treatment parameters, and clinical outcomes.

• Construct multi-center data-sharing platforms to facilitate developing and validating AI-driven predictive models across diverse patient populations.

• Establish real-time prediction and early warning systems to identify patients at risk of adverse events and guide proactive interventions.

The expansion of these research directions will not only enhance the precision of PRP therapy but also provide paradigmatic solutions for personalized treatment in the field of regenerative medicine. By integrating cutting-edge technologies from multiple disciplines and promoting the in-depth development of precision medicine in treating musculoskeletal diseases, we can ultimately achieve a paradigm shift from empirical medicine to precision medicine.

Conclusion

This study comprehensively evaluated the application of platelet-rich plasma (PRP) in treating knee osteoarthritis through systematic literature analysis and machine learning methods. The research findings reveal that the therapeutic effects of PRP are closely associated with several key parameters: a concentration range of 600–900 × 10⁹/L, an injection frequency of 3–5 times (at intervals of 7–14 days), and an early intervention strategy may yield better treatment outcomes.

Comparative analysis with traditional conservative treatment methods demonstrates the unique value of PRP therapy. In particular, PRP shows synergistic effects when combined with hyaluronic acid, and its favorable safety and tolerability profile provides crucial assurance for clinical application. The utilization of machine learning models offers new insights for the formulation of personalized treatment plans.

However, long-term follow-up data in existing studies remain insufficient, and the uniformity of evaluation criteria needs improvement. Future research should focus on these aspects and further explore the application of artificial intelligence techniques in clinical practice to provide patients with more precise and effective treatment options.

Author contribution

Author Contributions Statement Chengjing Wang Manuscript writing and revision Figure and table preparation Data processing and analysis Literature review and selection Bowen Yao Manuscript grammar review Structural and logical flow assessment Information extraction and verification Both authors have read and approved the final manuscript and confirm their individual contributions to the research. Authors' Responsibilities: Chengjing Wang took primary responsibility for the scientific content, research interpretation, and manuscript composition. Bowen Yao provided critical editorial support and ensured the manuscript's clarity and coherence. All authors have reviewed and agreed to the published version of the manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

No datasets were generated or analysed during the current study.

Ethical approval and consent to participate

Not applicable as this is a systematic review and meta-analysis.

Consent to publications

Not applicable as this is a systematic review and meta-analysis.

Declarations

Competing interests

The authors declare no competing interests.