Beyond DNA: An Explainable AI-Driven Multimodal Framework for Long-Term Missing Person Identification: A Systematic Review and Conceptual Model

Every year, thousands of people vanish, leaving behind a trail of silence that families and investigators struggle to fill. In the world of forensic science, the goal is to give a name to the unknown, a task that has traditionally relied on a few powerful tools. The most trusted of these is DNA profiling, which compares genetic material to find a match, but this method fails if no biological sample is available or if the reference sample from a relative is missing. Other methods, such as analyzing skeletal remains to estimate age and height or comparing facial photographs, offer clues but are often limited by the quality of the evidence or the passage of time. When a person has been missing for years, their face changes, bones may be damaged, and records can be incomplete, leaving investigators with fragments of information that do not fit together easily.

For decades, scientists have tried to solve these puzzles by looking at one type of evidence at a time, but the reality of a missing person case is rarely so simple. A single photograph might be too blurry, a DNA sample might be degraded, and a dental record might be lost. The challenge lies in weaving these disparate threads—visual, biological, and contextual—into a single, coherent picture. This is where the work of researchers at the National Forensic Sciences University comes in. They have developed a new way to approach these investigations by using artificial intelligence, not as a magic box that spits out an answer, but as a transparent partner that helps human experts make sense of the chaos.

The researchers built a system designed to handle the messy, incomplete nature of real-world missing person cases. Instead of relying on a single method, they created a framework that brings together different types of evidence: facial images, skeletal measurements, biological data, demographic details, and the circumstances surrounding the disappearance. In the past, computer programs used for identification often acted as "black boxes," providing a result without explaining how they reached it. This lack of transparency is a major problem in forensics, where a wrong identification can have life-altering legal consequences. To fix this, the team integrated a feature called "explainable artificial intelligence." This technology does not just guess an identity; it shows its work. It highlights exactly which pieces of evidence were most important in making a decision, allowing a human expert to see the logic behind the computer's suggestion and verify that it makes scientific sense.

To test their idea, the team constructed a massive digital dataset representing 1,200 missing person cases. These records were not perfect; they mimicked the real world where some cases had clear photos, others had DNA, and many had a mix of partial information. The researchers fed this data into their new system and compared its performance against older methods that looked at only one type of evidence at a time. The results were clear: the system that combined all the information was significantly better at identifying individuals than any single method on its own. When the system used only facial features, it achieved an AUC of 0.82. When it used only biological evidence, it achieved an AUC of 0.88. However, when the system fused all the available clues together, its AUC rose to 0.96. This improvement suggests that by letting different types of evidence support and fill in the gaps of one another, investigators can overcome the limitations of degraded or missing data.

Crucially, the study did not stop at just getting the right answer; it focused on how the answer was reached. The researchers used specific tools to break down the computer's decision-making process. They found that the system did not rely on just one factor, such as a face match, but weighed multiple factors together. For instance, the system might determine that a match was likely because the facial features were similar, the estimated age from the bones aligned with the missing person's timeline, and the geographic location of the find was consistent with the person's history. The system assigned a specific weight to each of these clues, showing that the facial image was the strongest indicator, contributing nearly 29 percent to the final decision, while other factors like geographic consistency and biological profile estimates played supporting roles. This level of detail allows a forensic expert to look at the computer's output and say, "I see why it made that connection," or "This part of the reasoning seems off," ensuring that the technology remains a tool for human judgment rather than a replacement for it.

The study also examined where the system failed to identify a person correctly. By analyzing these errors, the researchers discovered that the most common reasons for mistakes were poor-quality images, a complete lack of biological evidence, or conflicting information between different sources. This finding is vital because it tells investigators exactly where to focus their efforts. If a case lacks a clear photograph, the system knows to rely more heavily on other data, but if the evidence is too contradictory or of too low quality, the system signals that a match cannot be made with confidence. This honesty about uncertainty is a key part of the framework, preventing the false hope that can come from an overconfident computer guess.

The researchers emphasize that this system is not a final verdict but a decision-support tool. In the legal and forensic worlds, the responsibility for an identification must always remain with a human expert who can weigh the scientific evidence against the context of the case. The new framework simply provides a way to organize complex information and highlight the most promising leads. By making the artificial intelligence transparent, the researchers have created a bridge between advanced computing power and the rigorous standards of forensic science. The work suggests that in the future, solving long-term missing person cases will not depend on finding a single perfect piece of evidence, but on the ability to intelligently combine the many imperfect clues that are often available. This approach offers a path forward for cases that have remained cold for years, turning a scattered collection of data into a clear, understandable, and scientifically sound path toward identification.