Results

Study selection

The literature search identified a total of 269 records, including 148 records from PubMed/MEDLINE and 121 records from Scopus. After removing 80 duplicate records and one retracted article, 188 records remained for title and abstract screening. Of these, 118 records were excluded because they did not meet the eligibility criteria. The remaining 70 reports were assessed for full-text eligibility, of which two could not be retrieved.

Following full-text evaluation, four additional studies were excluded due to the absence of artificial intelligence or machine learning approaches, the use of simulated datasets without real postmortem data, or the lack of predictive modelling of the postmortem interval. Ultimately, 64 studies met the inclusion criteria and were included in the qualitative synthesis [11–74]. The study selection process is illustrated in the PRISMA flow diagram (Fig. 1).

Fig. 1: PRISMA Flow diagram of the systematic review

Characteristics of included studies

The main characteristics of the included studies are summarized in Supplementary Table S2. Overall, the 64 studies investigated a broad range of biological, molecular, imaging, and environmental datasets for artificial intelligence–based PMI estimation. Most investigations relied on experimental animal decomposition models, particularly rodents and pigs, although several studies employed human cadavers, skeletal remains, or retrospective forensic datasets.

The included studies evaluated both continuous PMI prediction and classification-based approaches. Investigated PMI windows ranged from early postmortem changes occurring within hours after death to long-term decomposition intervals extending over several months or years. The principal methodological categories identified in the literature are summarized in Table 1 and discussed in detail in the following sections.

AI approach | Biological source | Common AI/ML models | Number of studies | Typical PMI window | Reported performance
Microbiome-based models | Skin, gut, oral cavity, soil, bone, burial environments | Random Forest, XGBoost, Gradient Boosting, ANN | 29 | Hours → months | MAE ≈ 0.6 h – 5 days; R2 up to 0.95
Metabolomics/proteomics | Blood, muscle, liver, multi-organ tissues | SVR, neural networks, Random Forest, stacking ensembles | 11 | Hours → days | MAE ≈ 0.07–2 days; AUC up to 0.98
Spectroscopy-based approaches | Vitreous humor, skeletal muscle, bone, puparia | FTIR, ATR-FTIR, NIR + SVR/ANN/PLS models | 7 | Hours → months | RMSE ≈ 1–2 days (early PMI); classification accuracy up to 100%
Imaging-based AI (radiomics/pathology) | PMCT scans, histopathology slides, corneal imaging, hyperspectral imaging | CNN, ResNet, Vision Transformer, radiomics ML classifiers | 11 | Early PMI (hours–days) | Accuracy ≈ 78–96%; AUC ≈ 0.75–0.96
Forensic entomology models | Blowfly larvae, pupae, puparia | ANN, SVR, XGBoost regression | 6 | Days → weeks | Accuracy ≈ 80–94%; MAE ≈ 1–7 developmental units
Environmental/taphonomic ML datasets | Forensic decomposition databases, GIS environmental variables | Random Forest, machine learning classifiers | 2 | Days → months | R2 ≈ 0.82

Microbiome-based models

Microbiome-based approaches represented the largest group among the included studies, accounting for 29 of the 64 investigations. These models analyzed microbial succession in different anatomical and environmental contexts, including skin, gut, oral cavity, gravesoil, bone, burial environments, and aquatic decomposition settings.

Most studies employed Random Forest algorithms because of their ability to manage high-dimensional sequencing datasets and nonlinear biological interactions. Several investigations used murine or porcine experimental decomposition models, whereas others analyzed human cadavers or skeletal remains.

Human cadaver studies demonstrated promising predictive performance. Hu et al. analyzed intestinal microbiota from 63 cadavers and reported high predictive accuracy for PMI estimation [28], while Chen et al. investigated oral and nasal microbiota in 88 human swab samples and achieved mean absolute errors ranging from 2.16 to 5.14 days [18]. Other studies focused on environmental microbial communities associated with decomposition, including gravesoil bacteria, aquatic biofilms, and bone microbial succession.

Among experimental animal models, Liu et al. reported MAE values of approximately 20 h with coefficients of determination approaching R2 ≈ 0.95 in murine decomposition models [43]. Similarly, Zhao et al. demonstrated strong predictive performance using oral microbiome sequencing during rat decomposition, with R2 values approaching 0.94 [74].

Overall, microbiome-based models demonstrated some of the most consistent predictive performances across broad PMI ranges, although substantial variability remained related to environmental conditions, sample origin, and decomposition context.

Metabolomics and proteomics approaches

A total of 11 studies investigated metabolomics- and proteomics-based approaches for PMI estimation. These methods aimed to identify biochemical alterations associated with postmortem degradation processes, including metabolite accumulation, tissue biochemical changes, and protein fragmentation.

Most investigations analyzed blood, skeletal muscle, liver tissue, or multi-organ datasets using mass spectrometry–based analytical platforms combined with machine learning algorithms such as support vector regression, Random Forest, neural networks, and ensemble learning models.

Several studies demonstrated particularly promising performance for early PMI estimation. Magnusson et al. applied neural network models to large-scale human blood metabolomic datasets including more than 4,800 individuals and reported mean absolute errors of approximately 1.45 days [47]. Similarly, Li et al. developed a multi-omics stacking ensemble model integrating metabolomics, proteomics, and FTIR data, achieving AUC values up to 0.98 with very low prediction errors [38]. Fan et al. further demonstrated robust predictive performance across multiple temperature conditions using multi-organ GC–MS metabolomics datasets [25].

Proteomic approaches also showed encouraging results. Du et al. combined multi-organ proteomic fragments with ensemble machine learning methods and reported classification accuracies exceeding 0.89 in external validation cohorts [24]. Other studies investigated skeletal muscle or bone protein degradation patterns for late PMI estimation and demonstrated promising classification performance for long-term decomposition intervals.

Overall, metabolomics and proteomics approaches appeared particularly effective for early PMI estimation and multidimensional forensic datasets.

Imaging-based AI models

Imaging-based artificial intelligence approaches were reported in 11 studies and included postmortem computed tomography (PMCT), pathology whole-slide imaging, corneal opacity analysis, and hyperspectral imaging techniques.

Recent investigations increasingly employed deep learning architectures such as convolutional neural networks (CNNs), ResNet models, DenseNet, and Vision Transformers. Imaging-based models were primarily applied to early PMI estimation and interval classification tasks.

Ivanov et al. developed a 3D convolutional neural network based on PMCT imaging datasets obtained from 123 human cadavers and demonstrated promising classification performance for different PMI intervals [30]. Klontzas et al. used PMCT radiomics combined with XGBoost classifiers and reported AUC values of approximately 0.75 for distinguishing PMI intervals below and above 12 h [33].

Histopathology whole-slide imaging represented another emerging field. An et al. and Liang et al. applied deep learning architectures to pathology imaging datasets derived from human and animal decomposition models, achieving classification accuracies ranging from approximately 78% to over 93% [11, 12, 41, 42].

Additional imaging-based approaches included corneal opacity analysis and hyperspectral imaging of entomological specimens. Overall, imaging-based AI demonstrated promising performance, particularly in controlled experimental settings and early PMI estimation scenarios.

Spectroscopy-based approaches

Seven studies investigated spectroscopy-based approaches for PMI estimation, primarily using Fourier transform infrared spectroscopy (FTIR), attenuated total reflectance FTIR (ATR-FTIR), and near-infrared spectroscopy (NIR).

These methods analyzed vitreous humor, skeletal muscle, bone tissue, and insect puparia to detect biochemical alterations associated with decomposition. Artificial neural networks, support vector regression, partial least squares models, and Random Forest algorithms were commonly employed.

Several studies demonstrated excellent predictive performance during early PMI intervals. Zhang et al. reported ANN-based prediction errors of approximately 2 h using vitreous humor FTIR analysis [66], whereas Deng et al. achieved RMSE values of approximately 1.8 days using ATR-FTIR spectroscopy of human skeletal muscle [22]. Schmidt et al. also reported classification accuracies approaching 100% using NIR spectroscopy of human skeletal remains [53].

Other investigations explored long-term PMI estimation using weathered insect puparia or skeletal remains, suggesting that spectroscopy-based approaches may also provide value for extended decomposition intervals.

Forensic entomology and environmental/taphonomic models

Six studies focused on forensic entomology–based AI models, primarily analyzing blowfly larvae, pupae, puparia, or cuticular hydrocarbons for minimum PMI estimation. Artificial neural networks, support vector regression, and XGBoost regression models were commonly applied to insect developmental datasets.

These studies generally demonstrated high predictive performance, with classification accuracies ranging from approximately 80% to 94%. Several investigations also reported promising applications of hyperspectral imaging and ATR-FTIR spectroscopy for estimating developmental stages of forensically important insects.

Environmental and taphonomic machine learning datasets were less frequently investigated. Weisensee et al. developed a large forensic taphonomy database integrating GIS environmental variables and decomposition features from more than 2,500 forensic cases, achieving R2 values of approximately 0.82 [61]. Similarly, Korgesaar et al. demonstrated reliable PMI interval classification using environmental and taphonomic variables derived from retrospective forensic datasets [34]. Sharif et al. and Yu et al. further demonstrated promising performance of machine learning and deep learning models applied to puparial chemical and morphological analysis for PMImin estimation [54, 64].

Validation strategies and predictive performance

A wide range of machine learning algorithms were applied across the included studies. Random Forest represented the most frequently used approach, particularly in microbiome-based investigations, whereas deep learning architectures predominated in imaging-based studies.

Most studies relied on internal validation approaches, including cross-validation or train–test splits, while independent external validation cohorts were reported less frequently. Overall, AI-based models demonstrated promising predictive performance across different methodological categories, with reported MAE values ranging from a few hours to several days depending on the investigated PMI interval and biological matrix.

However, predictive performance varied substantially according to dataset size, environmental variability, decomposition conditions, and validation strategy. Studies employing external validation generally reported lower but more methodologically robust predictive performance compared with internally validated experimental datasets.

Risk of bias assessment

The risk-of-bias assessment demonstrated substantial methodological heterogeneity across the included studies. Most studies were classified as presenting moderate risk of bias, mainly due to limited external validation, relatively small experimental datasets, and the frequent use of animal decomposition models.

Only a minority of studies employed independent external validation cohorts, whereas most investigations relied on internal validation approaches such as cross-validation or train–test splits, potentially leading to optimistic estimates of predictive performance.

Studies based on human forensic datasets combined with external validation generally demonstrated lower methodological risk and greater generalizability. Conversely, studies using highly controlled experimental conditions or limited sample sizes showed increased concerns regarding reproducibility and real-world forensic applicability.

The detailed study-level assessment is reported in Supplementary Table S1.

Figure 2 illustrates the conceptual framework of AI-based approaches for postmortem interval estimation and the main categories of data sources and machine learning models identified in the literature.

Fig. 2: Conceptual framework of artificial intelligence–based approaches for postmortem interval estimation. Different types of postmortem data—including microbiome profiles, molecular biomarkers, imaging features, spectroscopic signals, entomological evidence, and environmental variables—can be integrated using machine learning and deep learning models to generate predictive estimates of the postmortem interval

Taken together, these findings suggest that AI-based PMI estimation models demonstrate substantial potential across multiple forensic domains, although methodological heterogeneity and limited external validation remain important limitations.