Various clinical trials of stem/progenitor cell transplantation for PAD or CLTI have identified several predictors and biomarkers of response to cellular therapy. Klepanec et al. [25] reported several parameters that delineated the R group from the NR group. Compared with the NR group, the R group exhibited a two-fold higher absolute CD34⁺ cell count but a similar number of total bone marrow mononuclear cells (BMNCs). Additionally, the CRP levels and total leucocyte counts in the R group were lower than those in the NR group. The general inflammation process and the number of CD34⁺ cells are directly associated with the cellular therapy outcomes [25,26]. For example, three independent clinical trials reported that increased levels of IL-6, CRP, serum leucocyte count, fibrinogen, and basic fibroblast growth factor are associated with a weak cellular therapy response. Two studies have reported that age < 50 years was an independent predictor of improved cellular therapy response in patients with CLTI [26,27]. Consistently, Jaiswal et al. reported that the frequency of somatic mutations markedly increased in patients aged > 50 years [5,28]. Moreover, patients with myelodysplastic syndrome (MDS) aged > 65 years and harboring TET2, DNMT3A, and AXSL1 mutations were susceptible to PAD and systemic inflammation [29]. These findings indicate a correlation between inflammation and CHIP, as well as CVDs, including PAD. PAD and coronary artery disease (CAD) are associated with decreased HSC proliferation and inhibition of angiogenesis, which lead to tissue ischemia and inflammation [15,30]. In the PERFECT trial, ML determined that the response of bone marrow stem cell transplantation-based myocardial regeneration was correlated with PLCG1, LPCAT2, AP1B1, AFAP1, GRB2, KLF8, and MARK3 expression levels and serum EPO and VEGF levels but was not correlated with the expression of CHIP-related genes (DNMT3a, TET2, and ASXL1). Sequencing data revealed that the R group exhibited 161 differentially expressed genes when compared with the NR group. Mutation analysis revealed that the number of specific variants in the R group (48 genes) was lower than that in the NR group (224 genes). Additionally, R-related and NR-related genes determined using ML were correlated with SH2B3, as well as with other regulatory genes, such as NOTCH2, PDCD1/PD-1, and CD133 [10].

Clinical studies have revealed that autologous G-CSF-mobilized peripheral blood EPCs exert higher therapeutic effects than BMMCs or PBMCs in patients with TAO. Arai et al. [32] demonstrated that the legs injected with CD34⁺ exhibited a significantly higher ankle-brachial index (ABI) (0.56 ± 0.04) than those injected with BMMCs (0.53 ± 0.06) at week 4 after cell therapy. Similar improvements in transcutaneous oxygen pressure (TcPO2) were observed in both CD34⁺-transplanted (from 27 ± 4 to 37 ± 3; p < 0.05) and BMT-transplanted groups (from 24 ± 6 to 32 ± 8; p < 0.05) at week 4 post-implantation. One clinical trial demonstrated that CD34⁺ cells dose-dependently prevented major or minor amputation when compared with the placebo. At month 6 post-injection, 67%, 43%, and 22% of the control, low-dose-administered, and high-dose-administered groups underwent a major or minor amputation (p > 0.137), respectively. This trend continued at month 12 post-injection, with 75%, 45%, and 22% of the control, low-dose-administered, and high-dose-administered groups undergoing amputation (p > 0.058), respectively. The amputation rate in the combined cell-treated groups was lower than in the control group (month 6, p > 0.125; month 12, p > 0.054). The low-dose-administered and high-dose-administered groups exhibited improved amputation-free survival at months 6 and 12 post-injection [24].

The long-term clinical outcomes of CD34⁺ transplantation revealed that compared with those at the baseline, toe-brachial pressure index (TBPI) and T_cPO₂ significantly improved at week 12 post-transplantation in both CLI and TAO cases but improved only at week 52 in TAO cases. The improvements in TBPI and T_cPO₂ were sustained until weeks 156 and 208, respectively. The ulcers completely healed in all patients with TAO and two patients with atherosclerotic PAD at week 52 [19]. A recent randomized, single-blinded trial evaluated the 12-month treatment outcomes of purified CD34⁺ cell transplantation and PBMC transplantation for advanced TAO. The total amputation rates at month 6 post-transplantation in the CD34⁺-transplanted and PBMC-transplanted groups were 28.0% and 16.0%, respectively (p > 0.343), which remained unchanged at month 12 post-transplantation. Furthermore, the complete wound-healing rates did not significantly differ between the groups at months 3, 6, and 12. The two groups exhibited significantly increased ABI, toe-brachial index (TBI), T_cPO₂, and pain-free walking time (PFWT) values over time when compared with baseline values (ΔABI, ΔTBI, ΔT_cPO₂, and ΔPFWT, respectively) [33]. Another clinical trial evaluated the long-term efficacy of autologous BMMC transplantation in ASO and TAO [34]. The four-year amputation-free rates in the control ASO, BMMC-transplanted ASO, control TAO, and BMMC-transplanted groups were 0%, 48%, 6%, and 95%, respectively. ABI and T_cPO₂ significantly increased after one month in the BMMC-transplanted TAO group and remained high during the three-year follow-up visit. In contrast, ABI and T_cO₂ significantly increased only in the first month in the BMMC-transplanted ASO group and gradually decreased during the three-year follow-up before finally returning to baseline levels [34]. The 10-year amputation-free survival rates in the autologous BMMC-transplanted and aspirin-treated groups were 85.3% (29/34) and 40% (6/15) (p < 0.0019), respectively. Autologous BMMC transplantation significantly decreased the ulcer area (p < 0.0001), TBI (p < 0.0001), T_cPO₂ (p < 0.0001), and pain score (p < 0.0001) [35]. Moreover, transplantation of autologous stem cells harboring mutations is reported to promote the expansion of mutant blood cell clones, leading to increased risks of CHIP-associated complications [36]. However, patients with TAO are younger than those with ASO. Thus, TAO cases exhibited decreased somatic mutations and inflammation in the peripheral blood.

These findings indicate that BMMC or PBMC transplantation is safe and feasible for patients with TAO. The parameters, such as the amputation-free rate (86%–95% in TAO and 40%–50% in ASO) and the enhanced ABI, TBI, T_cPO₂, and PFWT values compared with baseline values, indicated the effectiveness of cell therapies. BMMC or PBMC transplantation can exhibit similar efficacy as G-CSF-mobilized CD34⁺ cell transplantation in patients with TAO. However, purified or enriched transplantation of EPCs with a decreased somatic mutation profile must be performed for ASO. Thus, optimal stem cell-based therapy must be determined depending on the type of disease to reduce treatment expenditure.

3. AI-Based R and NR Stratification of Patients with PAD Undergoing Stem Cell Transplantation

AI has contributed to the analysis of clinical data in biomedical fields, including cardiovascular medicine. Additionally, AI assists and supports clinical decisions within a short duration [92].

3.1. High Data Quality Is Required for Accurate Prediction Models

The major challenges associated with clinical research on diseases (including PAD) include accessing and retrieving high-quality datasets [15]. The integrated analysis of high-throughput deep-sequencing data, patient phenotypes, images, and SNP loci enables the identification of robust biomarker candidates if controlled, manually curated, and high-quality data are used [93]. Data quality for advanced analysis is critical because it determines the outcomes of subsequent computational analysis [94]. Most studies utilize structured data for predictive modeling and ignore potentially valuable information in unstructured clinical notes, such as doctor reports. This is because of the challenges associated with integrating diverse reports, which are mostly handwritten text documents, into common AI algorithms. The integration of heterogeneous data types across electronic health records (EHRs) through deep learning (DL) techniques is reported to improve the performance of AI prediction models. Zhang et al. demonstrated that the models constructed based on the integration of unstructured clinical notes with structured data outperformed other models that utilize only unstructured notes or structured data [95]. Other DL methods have recently been reviewed for heterogeneous medical data [96] and image analysis [97]. However, AI-assisted clinical decisions should mainly be obtained from structured data to avoid overgeneralization from sparse datasets. Unstructured data allow adequate comprehensibility of the findings. Personalized therapy data, such as individual medical documents obtained from unstructured data, can be integrated. Thus, all data under investigation should be obtained under good practice (GxP) [98]. However, GxP cannot ensure the usage of appropriate or scientifically relevant methods nor the scientific significance of analyses or examinations. The data integrated into an AI study should be carefully analyzed, as using more data may decrease the signal-to-noise ratio [99].

3.2. Supporting Decision Making through AI-Based Complex Data Analysis

AI models are currently developed to mitigate the black-box effect, which leads to a lack of interpretability and transparency. Previously, this was the most important reason for the skeptical view of this technology held by patients and clinicians. This is due to the lack of trust in unfamiliar interfaces and hesitancy to rely on a machine or mathematical algorithm for making critical life decisions [100,101].

However, AI-based data analysis algorithms can be an independent extension of previously established statistical approaches for disease risk assessment, such as the recommendations by the American Heart Association/American College of Cardiology (ACC/AHA), which can predict the prognostic risk of CVD based on common risk factors, such as cholesterol, age, smoking, and diabetes [102]. However, several patients are not identified through the classical linear prediction models, and some patients are unnecessarily treated due to false-positive classifications [103,104]. Classical models may thus oversimplify complex, high-dimensional datasets using insufficient parameters. However, the integration of several parameters can result in overfitting the model. This so-called bias-variance dilemma occurs while generating an AI model and has a direct effect on the prediction accuracy, interpretability, and robustness to interpret new data. Complex models with many parameters and a high variance can often lead to overfitting. In these cases, the model adapts itself too closely to the training data and exhibits limited performance on new patients. In contrast, high-biased models that are not complex tend to ignore data points and important features, which ultimately leads to underfitting and decreased model accuracy [105]. This tradeoff must be specifically evaluated and taken into consideration for each dataset and newly trained model by testing different parametrizations. The increased availability of highly efficient AI algorithms has enabled the development of alternative approaches to classical linear prediction models. These models can utilize large, integrative datasets for improved prognosis and diagnosis [94]. In 2014, Dilsizian and Siegel reported that the vast amount of information obtained from patients and pre-clinical studies is too complex and heterogeneous for humans to comprehensively interpret without any technological support [106].

3.3. Identification of Responsive Patients for Cell Therapy

The number of patients undergoing genetic testing for various diseases, ranging from cancer to cardiomyopathy, has steadily increased. Additionally, RNA-seq can reveal specific mutational status and utilize the same guidelines used for DNA sequencing [107]. The ethical principles and history underlying clinical genetics will provide clinicians with improved tools to guide their practice and help patients navigate through complex medical-psychosocial terrain [108].

Recent biomedical studies have aimed to identify patient-specific biomarker signatures from high-throughput data to effectively predict postoperative results by stratifying patients into the R and NR groups before therapy. Wolfien et al. proposed a diagnostic strategy to predict the response of patients undergoing coronary artery bypass grafting to BMSC-mediated myocardial repair [10]. Predictive analyses in this case would provide useful insights to identify individuals who are most likely to benefit from BMSC treatment with patient-specific diagnostic characteristics.

This focused panel of molecular targets is consistent with the current comparison of AI approaches and traditional models for using administrative claims with EHR to predict heart-failure outcomes [109]. In this case, approaches with traditional logistic regression were compared in order to predict key outcomes in patients with heart failure, and the added value of predictive models was evaluated using EHR data. In total, 9502 patients (aged ≥ 65 years) with at least one heart-failure diagnosis were identified. Of these, 6113 were included in the training set, while 3389 were used as the testing set. The study comprised a large dataset with standard clinical parameters that do not have increased predictive value for ML-based stratification. The authors initially observed limited predictive capacity in the PERFECT study. Hence, they used additional molecular data, which improved the prediction accuracy of the ML model from 64% to 82% [15]. Further, the authors used specific molecular data, such as RNA-seq data. The stratification accuracy and sensitivity for a larger cohort increased by more than 10% [10]. Firouzi and Sussman suggested that this novel strategic approach of combining transcriptome profiling with precise patient phenotyping and AI-guided feature selection is a potentially valuable tool for advancing personalized medicine and cell-based therapies [110].

3.4. A New Diagnostic Tool Supporting Individual Disease Characterization

Generally, all studies utilize different ML methods to identify decision boundaries or specific patterns within patient data to individually characterize patients, identify similar groups or subgroups (referred to as clustering), or assign them to a certain disease, disease stage, or treatment option (referred to as stratification). The underlying “learning step” from data input to prediction involves a test of all available features for their capability of separating patients or subgroups in a supervised or unsupervised manner. The algorithm knows the correct label (e.g., control, treatment, and diagnosis) (the so-called ground-truth) in supervised ML. Each data point consists of manually selected features and the corresponding label (e.g., a data point represents a patient, the features are the clinical data, and the label is the stage of a specific disease). The algorithms attempt to identify a suitable relation between the features and the known label during the learning process, which is usually performed retrospectively on previously generated data. A trained ML model, which serves as a decision boundary, can now be applied to new patient data points to predict a label on previously unlabeled patient data. Here, a well-chosen patient parameter set without high imbalance ratios between the investigated groups and a well-defined ground truth is the key for accurate model performance. However, this can often not be assessed in medical data. New ML-based algorithms for dimensional reduction and visualization, such as t-SNE or UMAP, may enable new possibilities for classifying or reconsidering disease subgroups among patients with unknown labels. These unsupervised technologies for medical cases are predominantly used with heterogeneous data, especially for activated or aberrant signaling pathways (e.g., measured in peripheral blood or single-cell RNA-seq), genetic background (GWAS/TWAS), acquired SNPs, or other comorbidities [111,112]. Based on the integrative data, unique combinations and hybrid forms of diseases will be observed, as has been previously demonstrated on a smaller scale, in which AI will support their identification and characterization [10]. These investigations at this specific individual level would have been impossible without the use of AI. Therefore, these investigations assist in routine clinical decisions, especially concerning personalized cell therapy for PAD.

As previously mentioned, the analysis of medical images is one of the greatest success stories of AI, spanning the analysis of histopathological images [113], electrocardiograms [114], radiographs [115], magnetic resonance imaging slices [116], and many more [108]. Since imaging is very much an emerging field in PAD, this bears high potential because several already-established biomarkers could already be used for an unbiased patient stratification [117]. Similarly, a current review of Flores et al. points towards the infancy of AI in PAD but also foresees a broad spectrum of potential applications [118]. One of the first proofs of concept was implemented by Kim et al. [119] who used a deep convolutional neural network for the detection and assessment of the severity of PAD based on the analysis of brachial and ankle arterial pulse waveforms. These findings were compared with the state-of-the-art ankle-brachial index (ABI) using the virtual patients and showed, according to the authors, that DL may diagnose PAD more accurately and robustly than ABI. This work demonstrates a DL-based arterial pulse waveform analysis for affordable and convenient PAD screening, as well as the open challenges that need to be addressed for real-world clinical applications. However, one current limitation of imaging in PAD is the profound correlation of biomarkers with genetic alterations, which might be addressed with the help of AI, as already demonstrated in other fields, [120] as these ensemble models allow for complex data integration.

Nevertheless, a simple in silico identification of novel subgroups or biomarkers without a specific explanation for the choice a treatment option is insufficient. Therefore, AI algorithms in personalized PAD treatment also must utilize heterogeneous information at the gene level (e.g., GWAS, gene expression, pathway activity, identification of mutations) and phenotype level (e.g., cardiac functionality, angiogenesis potential, and inflammation status) to optimize predictions for cell therapy applications in the future. To achieve this, large patient cohorts are needed to utilize the full potential of personalized medicine therapies assisted by AI approaches.