The estimation of the age of the bloodstains represents a critical issue in forensic science, as it provides valuable temporal information for crime scene reconstruction. Previous studies have explored colorimetric features of blood degradation and applied statistical or kernel-based machine learning models to predict bloodstain age. In this work, we propose the application of a neural network regression model to analyze colorimetric parameters extracted from bloodstains monitored over a period of 24 hours. As shown by experimental results, the proposed approach reports reduced error compared to the state-of-the-art regression methods. Our findings demonstrate the potential of neural networks to capture nonlinear relationships in blood color degradation, ofering a promising and innovative framework for objective bloodstain age estimation.
https://www.docenti.unina.it/giovanni.acampora (G. Acampora); https://www.docenti.unina.it/autilia.vitiello (A. Vitiello)
CEUR Workshop
ISSN1613-0073
SVM with polynomial kernel are used with promising results. Instead, in [
Starting from this state of the art, this paper explores an alternative machine learning approach based on neural network regression for estimating bloodstain age. Neural networks are a particularly promising solution, as they can efectively capture subtle and nonlinear relationships within data, such as colorimetric features. In this study, the neural network architecture and its hyper-parameters were carefully tuned to optimize predictive performance. Experiments were conducted on a dataset of colorimetric parameters extracted from bloodstains monitored over a 24-hour period. The results demonstrate that the proposed approach outperforms existing methods in terms of standard regression evaluation metrics, highlighting its potential as a reliable tool for forensic bloodstain age estimation.
This section is devoted to describing materials and methods used in this work.
This study involves a dataset containing colorimetric features reported in [
Our study investigates the application of Neural Network Regression (NNR) to estimate the bloodstain age for forensic goals. The workflow of the study is reported in Fig. 1. Precisely, raw data have been split in training and test parts. Validation part has been extracted from training part for a tuning procedure. The optimized NN model have been used on the test part in order to analyse and discuss its performance. Hereafter, more details about our methodology are given.
The dataset was split into training and test sets following a typical 80–20% ratio. From the training portion, 10% of the samples were further set aside as validation data. Given the limited dataset size, no feature selection or extraction techniques were applied. At this stage, a correlation matrix (see Fig. 2) was analyzed to explore the degree of linear dependence between the features and the target variable. This preliminary assessment not only provides insights into the most influential predictors but also highlights the potential presence of more complex, non-linear interactions. Such evidence further supports the suitability of NNR, which is capable of modeling both linear and non-linear relationships.
Our approach consists of using a typical NN used for solving regression problem. Briefly, NNs are characterized by an architecture organized in layers: an input layer that receives the predictors, one or • Number of neurons per layer that controls the width of the network and the richness of the learned • Activation functions that determine the type of non-linearity introduced at each layer (e.g., rectifier function, logistic function and tangent function); • Learning rate that governs the step size during the optimization process and strongly impacts • Number of epoch that defines how many times the entire dataset is iterated during training; • Optimizer representing the algorithm used to update the weights (e.g., gradient descent and more hidden layers that transform the information through interconnected neurons, and an output layer that produces the final prediction. Each neuron computes a weighted sum of its inputs followed by a non-linear activation function, enabling the network to approximate complex functional mappings. Such a layered structure allows NNs to model intricate dependencies among features, making them particularly suitable when the data-generating process cannot be accurately described by simple linear models. During the training procedure, the choice of hyper-parameters of NNs strongly influences their performance. In particular, typical hyper-parameters are:
• Number of hidden layers that defines the depth of the network and its capacity to learn complex In our study, three key architectural hyper-parameters were tuned: the number of hidden layers, the number of neurons per layer, and the activation function. The optimal configuration was selected based on performance on validation data, after evaluating a total of 18 candidate models. Specifically, the number of hidden layers was varied within the set {1, 2, 3}, the number of neurons per layer within the set {32, 64}, and three activation functions were tested: rectifier (ReLU), logistic (sigmoid), and hyperbolic tangent (tanh). The best configuration resulted: 3 layers, 64 neurons and tanh as activation function. The optimizer was set to Adam, one of the most widely used and efective optimization algorithms. Training was performed for 1000 epochs, with the learning rate fixed at 0.001.
After finding the best hyper-parameters, the model was trained on the entire training set and tested on the mentioned test set. The final results were expressed in terms of standard evaluation metrics. The Mean Squared Error (MSE) measures the average squared diference between the predicted and actual functions; representations; convergence;
Adam). values: as: (1) (2) (3) (4) where and ̂ denote the true and predicted values, respectively. The MSE is non-negative (MSE ≥ 0), and lower values indicate better predictive performance. The Root Mean Squared Error (RMSE), defined provides an error measure in the same unit as the target variable, making it more interpretable. Hence, we will use RMSE instead of the simple MSE. RMSE is also non-negative, and smaller values correspond to better model accuracy. The Mean Absolute Error (MAE) computes the average absolute deviation: 1 =1 MSE = ∑ ( − ̂ ) ,
2 RMSE = √MSE,
1 =1 MAE =
∑ | − ̂ | , 2 = 1 − ∑ =1 ( − ̂ )2 ∑ =1 ( − )̄ 2 which is less sensitive to outliers compared to MSE. MAE is non-negative and lower values indicate better predictions. Finally, the Coeficient of Determination ( 2) evaluates the proportion of variance in the target variable explained by the model: where ̄ is the mean of the observed values. 2 ranges from −∞ to 1, with values closer to 1 indicating a better fit. A negative 2 indicates that the model performs worse than a simple mean predictor. Together, these metrics provide a comprehensive understanding of model accuracy and generalization.
This section is devoted to presenting the results of the application of the designed NNR for bloodstain
age estimation and a comparison of this approach to the state-of-the-art methods. In particular, the
comparison involves multiple linear regression (MLR), multiple quadratic regression (MQR), support
vector regression with gaussian kernel (SVMr), support vector regression with polynomial kernel
(SVMp) applied in [
As shown in Table 1, NNR outperforms the state-of-the-art methods by considering all evaluation metrics.
In this study, we proposed a NNR approach for estimating the age of bloodstains based on colorimetric parameters. The results demonstrate that the proposed method outperforms existing state-of-the-art techniques in terms of predictive accuracy, as evidenced by lower error metrics and higher 2 values on the test dataset.
Despite these promising results, several avenues for future work remain. First, the model could be further enhanced by incorporating additional features, such as environmental conditions (e.g., temperature, humidity, and surface type), which are known to influence bloodstain aging. Second, expanding the dataset with a larger number of samples and more diverse conditions would improve the network’s generalization capabilities. Finally, exploring more advanced architectures, including deep learning models with convolutional layers or attention mechanisms, could further improve predictive accuracy and robustness.
Overall, this work demonstrates the potential of neural network-based regression for forensic applications, ofering a reliable approach for estimating bloodstain age.
This study was funded by Ministero dell’Università e della Ricerca (MUR) of Italy in the context of project denoted as BLOODSTAIN in the program PRIN 2022 (grant number E53D23008040001).
The authors have not employed any Generative AI tools.