Deep Learning-Based STR Analysis for Missing Person Identification in Mass Casualty Incidents

Since its inception in the mid-1980s, the use of forensic DNA profiling has significantly increased [1, 2]. It is well-known as a technique used in forensic investigations for purposes such as identifying criminals from crime scene samples, determining paternity, and identifying the human remains of missing persons. It has become the gold standard for identifying victims, especially in mass-casualty incidents (MCIs) [3, 4].

MCIs can be defined as any sudden and unexpected events that have a negative impact on a community and cause more fatalities than the community is able to handle [5, 6]. There are several potential causes for MCIs, including natural disasters, armed conflicts, terrorist attacks, or accidents. After the Scandinavian Star ferry fire in 1990, DNA analysis was used for the first time to identify mass disaster victims.

For instance, on June 12, 2014, the Islamic State of Iraq and the Levant (ISIL) killed about 1700 members of the Iraqi army and security forces at Camp Speicher in Tikrit, Iraq, resulting in numerous casualties [7].

A forensic examination of human remains discovered in the aftermath of a war or other violent incident has two objectives. First, as part of criminal investigations, it is important to ascertain the cause and manner of death; second, it is important to identify any human remains and, if at all possible, return them to the victim's family. This paper takes the second objective into consideration.

The research proposes deep learning models for missing person identification in mass casualty incidents. rather than manual matching in the traditional approach that manually compares the human profiles of the missing person with those of his or her relatives. In manual matching, only three to four profiles can be compared simultaneously because it takes significant human effort to analyze 15 loci, each locus with two alleles for one short tandem repeat (STR) profile. Introducing an artificially intelligent system for missing person identification minimizes both effort and time consumption while maximizing the efficiency. The suggested system can also perform the identification regardless to the samples number. Familial dataset generation offers a good chance to create a familial database that may be useful in any other cases of missing person identification.

This paper introduces a deep learning modelling system to assess how closely human DNA profiles of missing persons and samples from living people are similar (mother, father, and brothers). The DNA-short tandem repeat (STR) profile, which consists of 15-loci with two alleles in each locus, is used as input data for this proposed deep learning system. The value that this system outputs indicates whether the missing person is a member of the reference family. The proposed system’s strength lies in the development and management of the dataset as well as in the optimal structuring of a deep learning model for the most accurate missing person identification. Figure 1 clarifies a clear overview of the proposed system.

1.jpg

Figure 1. Missing person identification

This paper is arranged as follows: Section 2 gives an overview of missing person identification based on DNA STR. Section 3 presents deep learning in general and gives a description of the three models used. Section 4 describes how the system was modelled and the simulations that were run. Section 5 presents a discussion of the results. Finally, the conclusions are given in Section 6.

This section gives an overview of deep learning and its application in the field of bioinformatics. Techniques that are used in this work; DNNs, GRUs, and Bi-GRUs are described as well.

Artificial intelligence (AI) is becoming increasingly important in forensic science as it offers the potential to increase the accuracy and effectiveness of various forensic tasks [20]. For example, AI can be used to analyse vast amounts of data in a relatively short length of time, automate tedious tasks, and identify patterns that humans might miss. However, it should be kept in mind that AI must be used in conjunction with human knowledge and judgment [21].

Deep learning is a branch of machine learning and AI that uses neural networks with multiple layers to analyze and model complex data patterns [22, 23].

In deep learning, large datasets are processed by neural networks, enabling them to autonomously learn patterns and relationships. This technique has been successfully applied across various bioinformatics applications to enhance classification accuracy. For instance, many studies have shown that deep learning outperforms traditional methods like random forest or support vector machine algorithms in predicting protein binding and accessibility to DNA sequences [24-28].

Using deep learning in missing persons identification through DNA matching has not been extensively studied. However, there are two relevant studies that apply artificial intelligence techniques to DNA analysis: Anggreainy et al. [29] utilize fuzzy inference for STR-DNA matching across 16 loci, and Siino and Sears [30] employ gradient descent logistic regression for kinship analysis based on 13 loci.

This paper explores and compares three deep learning techniques for an artificial intelligence system that uses DNA-STR data from 15 loci to identify victims: DNNs, GRUs, and bidirectional GRUs (Bi-GRUs). These techniques are detailed in Sections 3.1, 3.2, and 3.3.

3.1 DNNs

DNNs are a type of artificial neural network constructed from multiple layers of artificial neurons, or “units”. This approach was inspired by the way the human brain processes information [31].

The neural network has several layers, each of which processes and abstracts information in a unique way. The input data is processed at each layer before being forwarded to the one below it, with the final layer producing the output prediction or classification. When compared to other machine learning techniques, this enables the network to recognize high-level features and representations of the data, improving accuracy and performance [32].

To minimize the error between the predicted output and the true output DNNs are trained by adjusting the weights of the connections between units [33]. This process, known as backpropagation, is repeated multiple times until the error is reduced to an acceptable level. DNNs have proven effective at a variety of tasks, like speech and image recognition, machine translation, and natural language processing. They are particularly proficient at solving problems that call for the discovery of intricate connections between input and output data [34].

However, DNNs can be computationally expensive to train, and the selection of hyperparameters, such as the number of layers and the learning rate, can affect how they perform. They can also be prone to overfitting if the training data is not sufficiently large or diverse.

Each DNN neuron is mathematically represented in the formula below, where the output can be:

$Z=\sigma\left(\sum_t x_t \cdot W_t+b\right)$    (1)

where, x_t is the input vector, W_t is the weight vector, b is bias, Z is the output from the network, and σ is an activation function that may be a leaky rectified linear unit (Leaky ReLU) in all hidden layers or sigmoid in the output layer.

Leaky–ReLU:

$\sigma_l(x)=\max (a x, x)$    (2)

Sigmoid:

$\sigma_s(x)=\frac{1}{1+e^{-x}}$    (3)

3.2 GRUs

GRUs are a kind of RNN, and RNNs are a type of DNN specifically designed for processing sequential data [35]. RNNs are called “recurrent” because they have a loop structure that uses the output from one unit at a one-time step as input for the same unit at the following time step. The importance of processing sequential data lies in capturing temporal dependencies, understanding contextual information, modelling time series patterns, enabling natural language processing, supporting sequential decision-making, predicting future events, and facilitating sequential generation. These applications span a wide range of fields, highlighting the significance of effectively processing and analysing sequential data. RNNs offer these benefits by utilizing recurrent connections and memory mechanisms, allowing them to process sequential data, capture dependencies, and model temporal information effectively [36].

GRUs were firstly proposed by Cho et al. [37], it enables each recurrent unit to capture dependencies adaptively on different time scales. GRUs uses gating mechanisms in order to control the flow of information between the input, output, and hidden states [38-40].

The memory cell in a GRU is in charge of retaining and losing information over time. The memory cell is made up from two gates, an update gate and a reset gate [41]. Figure 3 depicts a memory cell that is ready for DNA applications.

3.jpg

Figure 3. GRU memory cell for DNA identification

A typical GRU unit consists of the following [35, 42]:

Reset gate: The reset gate control how much of the previous hidden state must be forgotten and how much of the current input must be incorporated into the updated hidden state.

$R=\sigma_s\left[\left(W_r \cdot X_t+V_r \cdot h_{(t-1)}\right)+b_r\right]$    (4)

where, R is the reset gate at time t; h_t−1 is the hidden state of the previous time step; X_t is the input at time t; Wr, V_r are the reset gate’s learnable weights for input and hidden state of reset gate and b_r is the bias; and σ_s is the sigmoid activation function.

After being transformed by the sigmoid function, the new values will all be between 0 and 1, which makes it possible for the gate to distinguish between unimportant and important data.

Update gate: The update gate determines how much of the previous hidden state should be retained and how much of the current input should be incorporated into the updated hidden state. The update gate is calculated next using the same formula as the reset gate; the only difference is the weights of the input vector and hidden state are distinct because each gate has its own set of weights, which implies that the final output vectors of each gate are different. As a result, the gates can be used for their particular purposes.

$U=\sigma_s\left[\left(W_u \cdot x_t+V_u \cdot h_{t-1}\right)+b_u\right]$    (5)

where U is the update gate at time t and W_u, V_u are the update gate’s learnable weights and b_u is the bias, for the update gate respectively.

Candidate hidden state: The candidate-hidden state is an intermediate value that is computed using the reset gate and the previous hidden state, as well as the current input. This process decides which information is preserved from the prior time steps alongside the new inputs.

$\tilde{h}_t=\boldsymbol{\sigma}_{\text {tanh }}\left[W_x \cdot x_t+\left(R \odot\left(W_h \cdot h_{(t-1)}\right)\right)+b_h\right]$    (6)

where, $\tilde{h}_t$ is the candidate hidden state at time t, σ_tanh is the hyperbolic tangent activation function, and ⊙ is the Hadamard (element-wise) product operator.

Hidden state: The final hidden state is a weighted combination of the previous hidden state and the candidate's hidden state, as determined by the update gate. If the update gate is close to 1, the candidate hidden state is mostly used to update the hidden state; if it is close to 0, the previous hidden state is mostly retained.

$h_t=(1-U) \odot h_{(t-1)}+U \odot \tilde{h}_t$    (7)

3.3 Bi-GRUs

Bi-GRUs are another kind of RNN that process input sequences in both forward and backward directions using GRUs [43, 44]. The motivation for using a bidirectional RNN is that the output at each time step may depend on the context from both past and future time steps.

A Bi-GRU processes input sequences using two separate RNNs, known as the forward and backward RNNs, which handle sequences in opposite directions [44]. Each time step's output from both RNNs is concatenated and serves as input for the subsequent step. This setup allows Bi-GRUs to integrate context from both past and future data points in the sequence. Bi-GRUs have proven effective in various applications, such as speech recognition, time series forecasting, and natural language processing. They are particularly valuable in tasks like language modeling and translation, where understanding context from both directions is crucial [45-48].

The Bi-GRU model is calculated based on the state of two GRUs, which are unidirectional in opposite directions. At time step t, the hidden state of the forward GRU is denoted as $\overrightarrow{h_t}$ while the backward GRU at time step t is $\overleftarrow{h_t}$.  

$\overrightarrow{h_t}=\operatorname{GRU}_{\text {forward }}\left(x_t, \overrightarrow{h_{t-1}}\right)$    (8)

$\overleftarrow{h_t}=\operatorname{GRU}_{\text {backward }}\left(x_t, \overleftarrow{h_{t+1}}\right)$    (9)

At each time step, the final result is obtained by concatenating the forward and backward hidden states:

$y_t=\overrightarrow{h_t} \oplus \overleftarrow{h_t}$    (10)

where, ⊕ indicates the process of concatenating two vectors.

The results were carried out based on Google-Colab, which is based on the Python programming language. The necessary libraries, including TensorFlow, Numpy, Pandas, Matplotlib, Scikit-learn, Seaborn, and Keras, were imported. The subsequent sections detail the partitioning of the dataset into training and testing sets, as well as the performance evaluations of the three proposed identification systems.

5.1 Training/testing dataset

During the training and testing phases, the entire dataset (151,580) was divided into 80% training and 20% testing sets. The training dataset was further divided into balanced labels, with half of the data labelled as 1 (indicating true children) and the other half labelled as 0 (indicating false children). Figure 10 shows the distribution of labels in a histogram. The training and testing processes were carried out separately for each of the proposed DNA deep learning models in order to make comparisons and evaluations.

10.jpg

Figure 10. Histogram of the training dataset

5.2 Performance comparison of the three classifiers

Applying deep learning models such as DNN, GRU, and Bi-GRU on DNA-STR datasets yields varying results in identifying missing persons.

The next step is the presentation of experimental results that support the work's primary goal, which is a binary classification for missing person identification in mass casualty incidents (MCI), the term MCI will be appended in each type of model results.

Accuracy results

In Figures 11 (a)-(c) accuracy and loss for DNN, GRU and Bi-GRU are presented respectively.

11a.jpg

(a) MCI-DNN model loss and accuracy

11b.jpg

(b) MCI-GRU model loss and accuracy

11c.jpg

(c) MCI-Bi-GRU model loss and accuracy

Figure 11. Performance of loss and accuracy for the three models

When comparing the performance of the three models over 10 epochs in Figure 11, it can be deduced that Bi-GRU and GRU models give the best performance in terms of accuracy and loss, while DNN has a higher loss and a lower accuracy value.

Confusion matrix results

Also, to compare the performance of the three models, the confusion matrix is also shown in Figures 12 (a)-(c) for the three models DNN, GRU, and Bi-GRU, respectively.

12a.jpg

(a) MCI-DNN confusion matrix

12b.jpg

(b) MCI-GRU confusion matrix

12c.jpg

(c) MCI-GRU confusion matrix

Figure 12. Confusion matrix for the three models

From the illustration of the confusion matrix, one can conclude that the Bi-GRU model gives the minimum total incorrect predictions (FP+FN) equal to 45, followed by GRU, which gives 53, and DNN, which gives the highest number of incorrect predictions equal to 2520. The TP and TN values are mostly equal due to equally distributed tested datasets.

ROC results

To support a more exhaustive analysis of the three binary classifier systems' performance, ROC is illustrated in Figures 13 (a)-(c) for DNN, GRU, and Bi-GRU, respectively.

13a.jpg

(a) MCI-DNN ROC

13b.jpg

(b) MCI-GRU ROC

13c.jpg

(c) MCI-Bi-GRU ROC

Figure 13. ROC curve for the three models

The ROC curve is used to assess system performance as a binary classifier and demonstrates strong results, as seen in Figure 13. It visualizes the trade-off between the true positive rate (TPR) and the false positive rate (FPR), with a perfect area under the curve equal to 0.999 for the Bi-GRU model, followed by 0.998 for GRU, and ending with 0.917 at DNN.

Tables 7 and 8 display a performance comparison of the three proposed models in terms of accuracy, loss, validation accuracy, and validation loss.

Table 7. Comparison between loss and validation loss for the three models

Table 8. Comparison between accuracy and validation accuracy for the tree models

Based on Table 7 provided, one can observe that the training and validation losses for the three models DNN, GRU, and Bi-GRU, trained for ten epochs. As the number of epochs increases, the training and validation losses for all models decrease. However, it can be observed that the Bi-GRU model achieves the lowest training and validation losses comparing with DNN and GRU models.

In terms of the DNN model, it can be seen that the validation loss decreases slower compared to the other models, indicating that the model may not be generalizing well to the validation data.

Precision, recall, and F1-score results

Table 9 appears to be a classification report showing the precision, recall, and F1-score for binary class (0 and 1) in a dataset of 30,316 instances. The metrics are compared among the three models (DNN, GRU, and Bi-GRU), and the DNN model has the lowest scores of the three models.

Table 9. Classification report

The DNN model appears to be the least effective of the three models, as it consistently achieves lower accuracy and validation accuracy scores compared to the GRU and Bi-GRU models, in which GRU and Bi-GRU are more effective in capturing the temporal dependencies and patterns present in the STR sequences, The GRU and Bi-GRU models outperform the DNN model because of their recurrent connections and memory mechanisms, which let them remember and apply data from earlier time steps. This feature is especially useful when processing sequential data, such as STR sequences, where the order and context of the data elements are critical for precise analysis. It's worth noting that the GRU model performs well, but not as well as the Bi-GRU model. This is likely because the Bi-GRU can process input sequences in both forward and backward directions, enabling it to capture more contextual information from the input data.

The average validation loss for the DNN model was found to be 0.2833, while the GRU model exhibited a significantly lower average validation loss of 0.01335 and the Bi-GRU model had the lowest average validation loss of 0.01021. A statistical analysis using a t-test revealed a highly significant difference (t-value = 11.0969, p < 0.001) between the mean validation losses of the DNN and GRU models, indicating that the GRU model outperformed the DNN model in terms of validation loss. Similarly, the Bi-GRU model also demonstrated superior performance compared to the DNN model. These findings suggest that the GRU and Bi-GRU models exhibit better predictive capability and are more effective in minimizing the validation loss compared to the DNN model.

Table 10 compares the proposed models with the baseline model regarding the number of profiles used and reported accuracy. It indicates that the proposed MCI-Bi-GRU model achieves the highest accuracy, utilizing 151,580 sample profiles.

Table 10. Models comparison with other related artificial intelligence works

To calculate the percentage of improvement, the accuracy of each proposed model can be compared with the baseline model [29]. It can be concluded that Bi-GRU improvement over Fuzzy is equal to 24.74%, GRU improvement over Fuzzy is equal to 24.55%, and DNN improvement over Fuzzy is equal to 14.68%.