ARTIFICIAL INTELLIGENCE AND DIFFERENTIAL
EQUATIONS IN POSTMORTEM INTERVAL ESTIMATION:
A CRITICAL REVIEW OF CADAVER COOLING MODELS
AND THEIR APPLICATIONS IN CRIMINAL
INVESTIGATION
INTELIGENCIA ARTIFICIAL Y ECUACIONES DIFERENCIALES EN LA
ESTIMACIÓN DEL INTERVALO POST MORTEM: REVISIÓN CRÍTICA DE LOS
MODELOS DE ENFRIAMIENTO CADAVÉRICO Y SUS APLICACIONES EN
CRIMINALÍSTICA
Ernesto Araujo Chavarro
CEEMA, Colombia
Yonathan Andrés Bermeo Orozco
CEEMA, Colombia
Cristian Eduardo Hernández Pillimue
CEEMA, Colombia
Laura Alejandra Gil Rinta
CEEMA, Colombia
Marian Cárdenas Poveda
CEEMA, Colombia
pág. 6004
DOI: https://doi.org/10.37811/cl_rcm.v10i5.26440
Artificial Intelligence and Differential Equations in Postmortem Interval
Estimation: A Critical Review of Cadaver Cooling Models and their
Applications in Criminal Investigation
Ernesto Araujo Chavarro1
e.araujo@udla.edu.co
https://orcid.org/0000-0002-7974-6969
CEEMA
Yonathan Andrés Bermeo Orozco
Yo.bermeo@udla.edu.co
https://orcid.org/0009-0008-1929-2863
CEEMA
Cristian Eduardo Hernández Pillimue
Cri.hernandez@udla.edu.co
https://orcid.org/0009-0004-6893-8940
CEEMA
Laura Alejandra Gil Rinta
la.gil@udla.edu.co
https://orcid.org/0009-0007-5314-2570
CEEMA
Marian Cárdenas Poveda
mari.cardenas@udla.edu.co
https://orcid.org/0009-0005-5298-7542
Laboratorio de Matemáticas y Física
ABSTRACT
This review examines the relationship between cadaver cooling, differential equations and artificial
intelligence in postmortem interval estimation, distinguishing computational plausibility from forensic
validation. A critical narrative review of 60 references combined documentary screening, individual
bibliographic identity checks and examination of available abstracts or full texts. The synthesis
distinguished controlled experiments, human observations and computational foundations. Thermal
models make mechanisms explicit, but their inversion depends on partly unknown initial and
environmental conditions. Machine learning incorporates metabolomic, molecular and imaging
information; heterogeneous performance measures do not establish overall superiority. Surrogate-based
optimization provides a documented integration of physical simulation and computation in human
cadavers, although validation remains limited. Physics-informed neural networks offer tools for inverse
problems; computational effectiveness alone does not establish forensic usefulness. Translation into
criminal investigation requires a defined application domain, separation of decedents between training
and evaluation, assessment of calibration and uncertainty, and external validation. A model should be
judged by the reliability of its inferences under documented conditions as well as its predictive fit.
Keywords: forensic medicine, artificial intelligence, models, theoretical, computer simulation
1 Autor principal
Correspondencia: e.araujo@udla.edu.co
pág. 6005
Inteligencia Artificial y Ecuaciones Diferenciales en la Estimación del
Intervalo Post Mortem: Revisión Crítica de los Modelos de Enfriamiento
Cadavérico y sus Aplicaciones en Criminalística
RESUMEN
Esta revisión analiza la articulación entre enfriamiento cadavérico, ecuaciones diferenciales e
inteligencia artificial en la estimación del intervalo post mortem, diferenciando plausibilidad
computacional y validación forense. Se desarrolló una revisión narrativa crítica de 60 referencias, con
depuración documental, comprobación individual de identidad bibliográfica y examen de resúmenes o
textos disponibles. La síntesis distinguió experimentación controlada, observaciones humanas y
fundamentos computacionales. Los modelos térmicos permiten explicitar mecanismos, pero su
inversión depende de condiciones iniciales y ambientales parcialmente desconocidas. El aprendizaje
automático incorpora información metabolómica, molecular e imagenológica; sus métricas resultan
heterogéneas y no permiten establecer una superioridad general. La optimización mediante modelos
sustitutos constituye una integración física-computacional documentada en cadáveres humanos, aunque
con validación limitada. Las redes informadas por física ofrecen herramientas para problemas inversos;
su eficacia computacional no acredita por sí misma utilidad pericial. El traslado a criminalística requiere
identificar el dominio de aplicación, separar cadáveres entre entrenamiento y evaluación, comprobar
calibración e incertidumbre y realizar validación externa. La contribución de un modelo debe juzgarse
por la confiabilidad de sus inferencias bajo condiciones documentadas, además de su ajuste predictivo.
Palabras clave: medicina legal, inteligencia artificial, modelos teóricos, simulación por computador,
modelos de aprendizaje predictivo
Artículo recibido 15 setiembre 2026
Aceptado para publicación: 21 octubre 2026
pág. 6006
Inteligência Artificial e Equações Diferenciais na Estimativa do Intervalo
Post Mortem: Revisão Crítica dos Modelos de Resfriamento Cadavérico e
Suas Aplicações na Criminalística
RESUMO
Esta revisão analisa a articulação entre resfriamento cadavérico, equações diferenciais e inteligência
artificial na estimativa do intervalo post mortem, distinguindo plausibilidade computacional e validação
forense. Foi realizada uma revisão narrativa crítica de 60 referências, com depuração documental,
verificação individual da identidade bibliográfica e exame dos resumos ou textos disponíveis. A síntese
diferenciou experimentos controlados, observações humanas e fundamentos computacionais. Os
modelos térmicos explicitam mecanismos, mas sua inversão depende de condições iniciais e ambientais
parcialmente desconhecidas. O aprendizado de máquina incorpora informações metabolômicas,
moleculares e de imagens; métricas heterogêneas não permitem estabelecer superioridade geral. A
otimização por modelos substitutos constitui uma integração físico-computacional documentada em
cadáveres humanos, embora com validação limitada. Redes informadas pela física oferecem
ferramentas para problemas inversos; sua eficácia computacional não comprova, por si só, utilidade
pericial. A aplicação à criminalística exige delimitar o domínio de uso, separar indivíduos entre
treinamento e avaliação, verificar calibração e incerteza e realizar validação externa. A contribuição de
um modelo deve ser avaliada pela confiabilidade das inferências sob condições documentadas, além do
ajuste preditivo.
Palavras-chave: medicina legal, inteligência artificial, modelos teóricos, simulação por computador,
modelos preditivos de aprendizagem
pág. 6007
INTRODUCTION
The postmortem interval (PMI) is the time elapsed between death and an explicitly defined reference
point, such as discovery, examination or sampling. The expressions postmortem interval, post-mortem
interval and time since death denote this interval in the literature, but they are interchangeable only
when they share the same endpoint. Confusing discovery with measurement changes the target variable
and may introduce an error independently of the algorithm used (Henssge & Madea, 2007; Wenzlow et
al., 2023; Wilk et al., 2022).
Thermal estimation reconstructs a history from the body’s current state. This inverse problem requires
separating observed information from assumed conditions: initial temperature, previous environmental
exposure, insulation and geometry. Henssge’s nomogram summarizes empirical relationships that are
useful under specified conditions. Numerical models represent heat transfer in greater detail, but require
parameters that cannot always be recovered at the scene. Greater physical detail does not remove
uncertainty about the body’s history (Henssge, 1988; Wilk et al., 2020; Wilk et al., 2022).
Machine learning shifts part of the effort from specifying mechanisms to identifying regularities in
thermal, molecular or imaging data. Available reviews report encouraging results alongside small
samples, heterogeneous validation and difficulties transferring models across settings. A network
identifying time categories in animal tissues and a model predicting days since death from human blood
address different questions and support different forensic uses (Bugelli et al., 2026; Sharma et al., 2022;
Sirago et al., 2026; Teixeira et al., 2025; Tournois et al., 2024).
This review examines what each model family contributes, what information it requires and what
evidence supports its application. Its central argument is that the relationship between physics and
artificial intelligence should be evaluated through identifiability, transportability and communication of
uncertainty. Hybrid models are useful when they integrate complementary information and enable their
assumptions to be tested; the label itself guarantees neither accuracy nor forensic validity (Collins et
al., 2024; Karniadakis et al., 2021; Kennedy & O’Hagan, 2001).
pág. 6008
METHODS
Review design and thematic axes
A critical narrative review combined bibliographic verification with methodological synthesis. Four
thematic axes were defined: cadaver cooling and thermometry; differential equations and inverse
problems; machine learning applied to PMI; and physical-computational integration, uncertainty and
validation. The final corpus comprises 60 distinct works. Classical contributions were retained when
they supported models or criteria still in use, together with relevant recent research. The reference count
describes the analytical corpus; it does not estimate the full extent of the literature.
This work is not presented as an exhaustive systematic review or meta-analysis. It had no prospective
registration, independent dual-reviewer screening or uniform access to full texts. PRISMA 2020
informed distinctions between identification, selection and reporting transparency, without attributing
unperformed procedures to the review. TRIPOD, TRIPOD+AI and PROBAST domains guided critical
reading; they did not replace a complete formal assessment of every study (Collins et al., 2015; Collins
et al., 2024; Page et al., 2021; Wolff et al., 2019).
Search strategy and study selection
References were consulted one by one through bibliographic records, accessible indexes and publisher
pages. Actual relevance was checked against the title and, where available, the abstract or full text. DOI,
authorship, title, year and publication details were compared. Crossref was used to verify bibliographic
identity, not to establish methodological quality. PubMed/MEDLINE, Europe PMC and publisher
pages supported content checks; targeted web searches located specific records. These checks closed
on September 25, 2026.
Information was extracted by examining available content to avoid confusing general artificial
intelligence or thermal-model literature with research applied to postmortem interval and its forensic
use. An existing DOI was insufficient to establish an attribution: mismatched titles and DOIs were
identified and corrected against the corresponding source. Duplicates were consolidated by DOI and
title comparison. Retracted records were excluded as substantive evidence, as were unrelated works
appearing in the initial compilation.
pág. 6009
The expanded strategy was organized as independent searches by thematic axis followed by combined
deduplication. The protocol covers Scopus, PubMed, Web of Science, IEEE Xplore, ScienceDirect,
Google Scholar, SciELO and Redalyc. Planned coverage must be distinguished from searches actually
performed: auditable exports of complete searches across all these platforms were unavailable, and their
execution is not claimed. In particular, including a journal hosted on ScienceDirect or indexed in Scopus
does not establish that an independent search of that database was conducted. Section 5.2.1 provides
complete search strings prepared for replication; retrieval counts remain undetermined.
Search strings and replication criteria
The following strings constitute a prepared protocol for extending and reproducing the search; no
retrieval counts or complete execution across all platforms are claimed. Date filters are initially omitted
to retrieve foundational work; recent literature is subsequently prioritized and classical exceptions
documented. Results from each axis should be exported separately, combined by union and
deduplicated by DOI and title. Requiring intersection of every axis is avoided because it would exclude
relevant physical foundations and AI studies that do not mention the other model family.
Scopus — thermal axis
TITLE-ABS-KEY(("postmortem interval" OR "post-mortem interval" OR "time since death" OR "cad
aver cooling" OR "postmortem cooling") AND (Henssge OR Newton OR "heat transfer" OR "differen
tial equation" OR "differential equations" OR "finite element" OR "finite difference"))
Scopus — artificial intelligence axis
TITLE-ABS-KEY(("postmortem interval" OR "post-mortem interval" OR "time since death") AND ("
artificial intelligence" OR "machine learning" OR "deep learning" OR "neural network" OR "neural ne
tworks" OR "random forest"))
PubMed — forensic application
("postmortem interval"[Title/Abstract] OR "post-mortem interval"[Title/Abstract] OR "time since deat
h"[Title/Abstract]) AND ("Artificial Intelligence"[Mesh] OR "Machine Learning"[Mesh] OR "machin
e learning"[Title/Abstract] OR "deep learning"[Title/Abstract] OR Henssge[Title/Abstract] OR "body
cooling"[Title/Abstract] OR "heat transfer"[Title/Abstract])
pág. 6010
Web of Science — integration
TS=(("postmortem interval" OR "post-mortem interval" OR "time since death") AND ("heat transfer"
OR Henssge OR "differential equation*" OR "machine learning" OR "deep learning" OR "physics-inf
ormed" OR "surrogate model*"))
IEEE Xplore — all metadata search
(("All Metadata":"postmortem interval" OR "All Metadata":"time since death") AND ("All Metadata"
:"machine learning" OR "All Metadata":"neural networks" OR "All Metadata":"heat transfer" OR "All
Metadata":"physics-informed"))
ScienceDirect — title, abstract or keywords
("postmortem interval" OR "time since death") AND ("body cooling" OR Henssge OR "machine learn
ing" OR "physics-informed")
Hybrid foundations — independent IEEE Xplore or ScienceDirect search
("physics-informed neural networks" OR "physics-informed machine learning" OR "surrogate model"
) AND ("heat transfer" OR "inverse problem" OR "uncertainty quantification")
Google Scholar — run the three queries separately
"postmortem interval" "machine learning"
"time since death" "body cooling"
"postmortem interval" "physics-informed"
SciELO and Redalyc — general fields with language variants
("intervalo post mortem" OR "intervalo postmortem" OR "tiempo de muerte" OR "tempo de morte") A
ND ("inteligencia artificial" OR "inteligência artificial" OR "enfriamiento cadavérico" OR "resfriamen
to cadavérico" OR "ecuaciones diferenciales" OR "equações diferenciais")
If an interface does not support nested Boolean operators, phrase combinations should be executed and
logged as independent queries. Free-text search terms are not presented as controlled descriptors. The
article’s five DeCS/MeSH descriptors retain conceptual identity across languages; identifiers and links
are provided in the traceability supplement.
pág. 6011
Inclusion and exclusion criteria
The review included primary studies and reviews on PMI estimation, cooling research with cadaver-
related relevance, and physical or statistical foundations needed to interpret hybrid models. Nonforensic
methodological literature was explicitly classified as transferable foundational work and was never
treated as medicolegal validation. English-language and other publications were considered when their
identity and contribution could be established through available abstracts, metadata or text.
Duplicates, retracted records, unresolved bibliographic inconsistencies and studies connected only
through generic terms such as temperature or artificial intelligence were excluded. Species
identification, cause-of-death diagnosis and injury classification without a relevant temporal outcome
were not considered equivalent to PMI estimation. Reviews contextualized the field and informed
methodological comparison; their constituent studies were not counted as independent observations in
addition to primary studies already included (Bugelli et al., 2026; Sharma et al., 2022; Tournois et al.,
2024).
Data extraction and critical appraisal
Extracted information included species, experimental unit, number of individuals when available, time
window, environment, input signals, model family, validation strategy, metrics and scope of
conclusions. Three evidence levels were distinguished: A, laboratory studies, animals or human
cadavers under controlled conditions; B, human observations related to medicolegal investigations,
which do not automatically imply external validation; and C, mathematical or computational evidence
without direct forensic application. Reviews and guidelines were classified as secondary or
methodological evidence.
Critical reading addressed five questions: whether the reference PMI was known or estimated; whether
observations from the same individual could cross dataset partitions; whether metrics described
regression or classification; whether calibration was assessed; and whether the validation environment
represented intended use. Missing information was treated as a limitation of accessible reporting, not
proof that a procedure had not been performed. No aggregate quality scores or pooled effects were
calculated because outcomes, units and designs were insufficiently homogeneous (Collins et al., 2024;
Riley et al., 2019; Wolff et al., 2019).
pág. 6012
RESULTS AND CRITICAL DISCUSSION
Overview of postmortem interval estimation
PMI is not directly observable in most retrospective investigations. Estimation links biological and
physical transformations to a timescale and depends on documenting modifying factors. The established
recommendation to integrate methods remains relevant: thermal signals, supravital muscle responses
and molecular markers may provide different constraints. Combining them requires understanding their
dependence; accumulating correlated indicators does not amount to assembling independent evidence
(Henssge et al., 1988; Henssge & Madea, 2007; Stigter et al., 2024).
Cooling is particularly informative while a thermal contrast with the environment persists. Molecular
degradation, microbial succession and postmortem imaging investigate other processes and temporal
windows. Reviews by Wenzlow and colleagues and Cianci and colleagues show that species, tissue and
storage conditions delimit marker interpretation. Changing modality does not remove environmental
effects; it changes how those effects influence the signal (Cianci et al., 2024; Wenzlow et al., 2023;
Yuan et al., 2023).
Responsible comparison must hold the prediction problem constant. A study of the first hours after
death cannot be ranked against a study of weeks by its raw error value; classification accuracy cannot
be translated into minutes either. Henssge and Madea emphasized that methods should be compared
over the same interval while accounting for variation in influential factors. This requirement becomes
especially important when technologies differ in capabilities and costs (Bugelli et al., 2026; Henssge &
Madea, 2007; Sirago et al., 2026).
Cadaver cooling and heat-transfer models
Classical approaches: Newton, Henssge and variants
A lumped Newtonian cooling representation assumes an effectively uniform body temperature, a
constant cooling coefficient and constant ambient temperature. As an idealization of heat exchange, it
is written as:
𝑑𝑇
𝑑𝑡 = −𝑘(𝑇 − 𝑇𝑎), 𝑇(0) = 𝑇0, 𝑘 > 0. (1)
pág. 6013
Its solution and temporal inversion are:
𝑇(𝑡) = 𝑇𝑎 + (𝑇0 − 𝑇𝑎)𝑒−𝑘𝑡, 𝑡 = 1
𝑘 ln (𝑇0 − 𝑇𝑎
𝑇 − 𝑇𝑎
). (2)
Here, 𝑇 denotes modeled temperature, 𝑇𝑎 ambient temperature, 𝑇0 initial temperature and 𝑘 an effective
rate with inverse-time units. The equation explains the inverse reasoning but does not itself validate a
cadaver estimate. Core-to-surface gradients and the initial cooling phase motivate more complex
representations. Rectal temperature should not be equated without justification to the uniform
temperature of a lumped body (Henssge, 1988; Henssge & Madea, 2007; Wilk et al., 2020).
In Henssge’s approach, the dimensionless ratio 𝑄 = (𝑇𝑟 − 𝑇𝑎)/(𝑇0 − 𝑇𝑎) is related to time through a
calibrated biexponential function. A classical branch for lower ambient temperatures uses:
𝑄(𝑡) = 1.25𝑒−𝑘𝑡 − 0.25𝑒−5𝑘𝑡. (3)
This expression illustrates an initially zero slope; it does not represent every branch or correction of the
method. Application requires rectal and ambient temperatures, body mass and correction factors for
cooling conditions; an initial temperature of 37.2 °C is commonly assumed. The appropriate
parameterization and application limits must be used rather than freely fitting 𝑘 while retaining the
designation of a validated method (Henssge, 1988; Otatsume et al., 2024).
Otatsume and colleagues implemented numerical solution using Solver and tabulated searches in Excel.
Nedugov proposed improvements in equation solving and error handling. These contributions reduce
reading, discretization or calculation problems, but a more accurate numerical root does not correct an
incorrect initial temperature or an unknown environmental history. The accuracy of the equation-
solving algorithm must be separated from the validity of the model represented by that equation (G. V.
Nedugov, 2021b; Otatsume et al., 2024).
6.2.2. Differential-equation models and physical assumptions
When ambient temperature varies, the lumped equation becomes 𝑑𝑇/𝑑𝑡 = −𝑘[𝑇 − 𝑇𝑎(𝑡)]. With
constant 𝑘, its solution is:
𝑇(𝑡) = 𝑇0𝑒−𝑘𝑡 + 𝑘 ∫ 𝑒−𝑘(𝑡−𝑠)
𝑡
0
𝑇𝑎(𝑠) 𝑑𝑠. (4)
pág. 6014
This identity, derived with an integrating factor, shows that present temperature depends on the
environmental trajectory rather than only its current value. Models with linearly varying ambient
temperature and dual-thermometry strategies attempt to represent part of that dependence. The linear
assumption must be checked against actual exposure; transfer into refrigeration introduces a regime
change that need not follow a single trend (G. V. Nedugov, 2021a; V. G. Nedugov & Nedugov, 2024,
2025).
Distributed conduction represents position 𝐱, heterogeneity and boundary conditions:
𝜌𝑐 ∂𝑇
∂𝑡 = ∇ ⋅ (𝜆∇𝑇) + 𝑞, 𝐱 ∈ 𝛺. (5)
Density 𝜌, specific heat 𝑐 and conductivity 𝜆 may differ between tissues. The term 𝑞 collects volumetric
sources requiring justification. One possible balance on an exposed surface is:
−𝜆∇𝑇 ⋅ 𝐧 = ℎ(𝑇 − 𝑇𝑎) + 𝜀𝜎(𝑇4 − 𝑇rad
4 ) − 𝑞solar. (6)
Radiative temperatures are absolute, in kelvin; ℎ is the convective coefficient, 𝜀 emissivity, 𝜎 the
Stefan–Boltzmann constant and 𝑞solar absorbed solar radiation. Ground or surface contact requires an
additional conduction or thermal-resistance condition. Equations (5)–(6) summarize modeling
principles; the terms retained depend on the scenario (Belykh & Vavilov, 2025; Wilk et al., 2020; Wilk
et al., 2022).
Pennes’ bioheat model was developed for living tissue and includes perfusion-related exchange. After
death, steady physiological perfusion should not be retained as though circulation continued. A cadaver
adaptation must justify which terms vanish, which persist transiently and how the initial temperature
distribution is described. Citing a bioheat equation does not establish that its parameterization is valid
in postmortem tissue (Pennes, 1948; Wilk et al., 2020).
Numerical models and limits to generalization
Finite differences and finite elements accommodate geometry, layers and boundaries omitted by lumped
models. Wilk and colleagues combined surface thermometry with finite-difference simulation in human
cadavers. This evidence supports experimental feasibility, but generalization requires scrutiny of
anatomical representation and the properties assigned to tissues, clothing and surroundings. More
pág. 6015
detailed geometry may also introduce poorly identifiable parameters (Wilk et al., 2020; Wilk et al.,
2022).
The history of changing environmental conditions provides an instructive comparison. Althaus and
Henssge investigated abrupt changes using objects with body-like cooling. Bovenschen and colleagues
evaluated back-calculation in 25 deceased individuals moved into refrigerated storage and found
substantial differences between reconstructed and observed temperatures. Transfer from experimental
objects to human cadavers therefore requires its own testing. The availability of a mathematical solution
does not justify assuming that it reproduces an actual thermal transition (Althaus & Henssge, 1999;
Bovenschen et al., 2021).
Solar exposure likewise cannot always be represented as a simple increase in air temperature. Belykh
and Vavilov studied substitute objects and cadavers exposed to radiation, incorporating a solar factor
dependent on material and environmental conditions. Their contribution identifies a boundary effect
that can alter the balance; reported errors should not be interpreted as guarantees for every climate,
fabric or surface. Evaluations must distinguish the data used to fit that factor from those used to test it
(Belykh & Vavilov, 2025; Kennedy & O’Hagan, 2001).
Artificial intelligence applied to PMI
Supervised learning and neural networks
Supervised learning fits a relationship between predictors and a reference PMI or temporal category.
Random forests combine trees to model nonlinear relationships, while deep networks can extract
representations from images or high-dimensional profiles. Architectural complexity does not correct
imprecise temporal labels or increase the number of independent individuals. Reviews should therefore
be read with attention to the biological unit and dataset separation protocol (Breiman, 2001; Sharma et
al., 2022; Sirago et al., 2026).
Løber and colleagues developed metabolomic models in rat tissues and tested them in an independent
animal dataset. An and colleagues used 12 pigs and added four for external validation of a multiorgan
pathomics strategy. Their reported 87.5% external accuracy concerns classification in that design; it
does not mean that time of death is estimated with a 12.5% error. Testing new individuals is an advance,
although transfer to humans remains unresolved (An et al., 2025; Løber et al., 2025).
pág. 6016
Human evidence involves other timescales. Magnusson and colleagues analyzed femoral blood from
4,876 individuals with intervals of 1–67 days. They reported mean and median absolute errors of 1.45
and 1.03 days in testing and 1.78 and 1.29 days in 512 individuals from another year and analytical
platform. Temporal and platform validation informs transportability, but does not make the model a
tool for resolving differences of a few hours in early PMI (Magnusson et al., 2026).
In imaging, Klontzas and colleagues included 97 human CT examinations and classified intervals on
either side of a 12-hour threshold, reporting an AUC of 0.75 for the combined model. Ivanov and
colleagues analyzed 544 examinations from 123 decedents. Distinguishing scans from individuals is
essential: repeated observations extend the measured trajectory, but are not independent human
replicates. Results depend on how these trajectories are separated during training and evaluation
(Collins et al., 2024; Ivanov et al., 2025; Klontzas et al., 2023).
Input variables: temperature, environmental context and uncertainty
Temperature acts both as a physical observation and as a modifier of degradation. Fan and colleagues
integrated multiorgan metabolomics and ambient temperature in rats; Ren and colleagues investigated
microbial succession in bromadiolone-exposed rats. These studies show that tissue, environment and
toxicants can alter the relationship between signal and time. Including temperature as a statistical
predictor may be useful, but is not equivalent to imposing conservation of energy (Fan et al., 2025; Ren
et al., 2026).
Gravesoil microbiota reflect interactions between the body and its ecosystem. Work by Cui and by
Metcalf supports investigating succession as a temporal signal; it does not establish microbial
composition as a universal clock. Substrate, sampling and decomposition conditions define the
inference domain. A model trained in one site–season combination may recognize environmental
characteristics rather than a transportable temporal relationship (Cui et al., 2023; Metcalf et al., 2016;
Yuan et al., 2023).
RNA profiles add variation associated with sample integrity, normalization and reference-gene stability.
Wang and colleagues investigated endogenous and exogenous signals in a mouse model. Cianci and
colleagues examined nine refrigerated porcine hearts using leave-one-heart-out validation and found
pág. 6017
that data representation affected prediction. Analytical reproducibility and biological validation must
be assessed separately (Cianci et al., 2024; Cianci et al., 2026; Y. Wang et al., 2026).
Performance metrics and calibration
For regression, mean absolute error and root mean squared error are:
MAE = 1
𝑛 ∑|𝑡̂𝑖 − 𝑡𝑖|
𝑛
𝑖=1
, RMSE = √1
𝑛 ∑(𝑡̂𝑖 − 𝑡𝑖)2
𝑛
𝑖=1
. (7)
Both require reporting units and the distribution of PMI. Mean signed error may approach zero because
overestimates and underestimates cancel. Correlation does not measure agreement either; residuals and
limits of agreement should therefore be examined while respecting repeated-measure dependence and
statistical assumptions (Bland & Altman, 1986; Sullivan, 2015).
AUC describes discrimination between classes, not the temporal distance of a prediction. Shen and
colleagues reported accuracy and AUC for animal CT models; Liang and colleagues distinguished
image-patch and slide-level results. Li and colleagues combined semi-supervised learning with
classifier selection. Their metrics must retain the analytical level and task for which they were obtained
(Li et al., 2026; Liang et al., 2025; Shen et al., 2024).
The review by Cuba-Gutierrez and colleagues reports a pooled AUC of 0.94 with an interval of 0.81–
1.08. Its upper bound exceeds the domain of AUC. This requires checking the interval-estimation
method and study compatibility rather than interpreting 1.08 as an attainable discrimination capacity.
The example shows why meta-analysis does not remove scale problems or replace assessment of its
assumptions (Cuba-Gutierrez et al., 2026; Gneiting & Raftery, 2007).
Calibration examines correspondence between predictions and observed frequencies or magnitudes.
Classification probabilities require calibration curves; temporal estimates benefit from examination of
intercepts, slopes and residuals, alongside prediction intervals. For a nominal 95% interval, coverage
should be assessed in new individuals together with interval width. Very wide intervals may achieve
coverage but offer limited usefulness; narrow intervals without adequate coverage communicate
unjustified certainty (Collins et al., 2024; Gneiting & Raftery, 2007; Van Calster et al., 2019).
pág. 6018
Hybrid approaches: artificial intelligence and physics
Effective integration versus parallel learning
A hybrid model explicitly integrates a physical mechanism with learned components or coupled
computational inference. It may learn equation parameters, construct a surrogate to accelerate solution
or penalize physical violations during training. Simply including temperature among dataset columns
does not qualify. Sparse identification of dynamical systems offers another route by selecting equation
terms from data, although recovering a parsimonious structure does not itself establish causality or
cadaver validity (Brunton et al., 2016; Karniadakis et al., 2021).
Wilk and colleagues coupled a thermodynamic simulation to optimization using a surrogate of the cost
function. The surrogate was constructed with radial-basis interpolation to explore perimortem and
ambient temperatures. This integration was documented in four human cadavers and should not be
described as a PINN. Mean signed error was 0.18 hours, with a standard deviation of 0.77 hours;
experimental sample size and dependence between measurement locations limit inference about
population performance (Wilk et al., 2022).
In a physics-informed neural network (PINN), an approximation 𝑇𝜃(𝐱, 𝑡) can be trained using:
ℒ = 𝜆𝑑ℒdata + 𝜆𝑓ℒPDE + 𝜆𝑏ℒboundary + 𝜆0ℒinitial. (8)
A physical residual can be defined as 𝑟𝜃 = 𝜌𝑐 ∂𝑡𝑇𝜃 − ∇ ⋅ (𝜆∇𝑇𝜃) − 𝑞. Raissi and colleagues established
this formulation for forward and inverse problems. Cai and colleagues investigated heat-transfer
applications, while Lu and colleagues developed tools for solving differential equations with networks.
These are level-C computational contributions, not trials of PMI estimation in criminal investigations
(Cai et al., 2021; Lu et al., 2021; Raissi et al., 2019).
Interpretability, physical guarantees and bias
A soft physical penalty does not guarantee exact satisfaction of the equation, its boundaries or integral
conservation. Wang and colleagues demonstrated gradient pathologies during PINN training. Before
interpreting estimated parameters, scaling, units, collocation-point distribution, residual errors and
initialization sensitivity should be assessed. A low aggregate loss may conceal an inadequately fitted
physical component (Karniadakis et al., 2021; S. Wang et al., 2021).
pág. 6019
Bayesian and adversarial formulations can represent uncertainty in solutions or parameters under
explicit assumptions. That uncertainty remains conditional on the chosen model, noise distribution and
available observations. If a model omits relocation of the body or a heat source, its posterior may be
narrow around an incorrect explanation. Discrepancy between simulator and reality must be considered
separately rather than automatically absorbed into parameters interpreted physiologically (Kennedy &
O’Hagan, 2001; L. Yang et al., 2021; Y. Yang & Perdikaris, 2019).
Explanations of tree-based models may identify variables contributing to a prediction. Their diagnostic
usefulness does not imply that those variables cause the process or that a local explanation validates
extrapolation. Rudin’s discussion of interpretable models is relevant to high-stakes decisions:
understandable alternatives should be compared with complex models under the same evaluation. The
burden of justification increases when additional predictive benefit is small and reconstruction of the
reasoning is difficult (Lundberg et al., 2020; Rudin, 2019).
Forensic validation and translation into criminal investigation
Laboratory studies and real-world settings
Validation should be described by the change it tests. A new animal dataset examines experimental
reproducibility; controlled human cadavers approximate relevant species-specific properties;
retrospective human cases evaluate aspects of practice; and an independent prospective evaluation
examines the entire procedure. No level automatically replaces the next. Reviews by Sharma and Sirago
demonstrate why preclinical and human evidence should remain distinct (Sharma et al., 2022; Sirago
et al., 2026).
Knowledge of time of death creates an additional difficulty. A label derived from witness accounts,
clinical records or a last-seen-alive window may carry uncertainty. Training against a previous Henssge
estimate and then presenting agreement with Henssge as independent validation would be circular. The
origin of every label should be documented; interval-valued labels should retain that structure in
evaluation instead of being replaced by apparently exact points (Collins et al., 2024; Henssge & Madea,
2007; Wolff et al., 2019).
Partitioning must occur by individual before any operation capable of transferring information between
datasets. Feature selection, imputation, normalization and hyperparameter tuning belong within
pág. 6020
training. Problems identified by Roberts and colleagues in clinical imaging provide a transferable
methodological warning about bias and deficient validation; that review is not evidence of PMI
performance. Its inclusion is justified exclusively by this analytical contribution (Collins et al., 2024;
Roberts et al., 2021).
Robustness to climate, clothing, body position and substrate
Temperature measured at discovery does not necessarily represent prior exposure. Refrigeration, solar
radiation, wet clothing and substrate contact alter heat exchange. Robustness tests should vary these
conditions in documented ways and assess both bias and coverage. Back-calculation and surface-
modeling results show that environmental history forms part of the inference problem (Belykh &
Vavilov, 2025; Bovenschen et al., 2021; Wilk et al., 2020).
Transfer to tropical environments, including Amazonian settings, requires local data; this corpus does
not establish operational validation in Caquetá. Thermal cycles, radiation, wet surfaces and indoor–
outdoor transitions should be evaluated through controlled protocols followed by independent human
cases. This proposal follows from boundary mechanisms and transportability limitations, not from
demonstrated regional superiority. Adding “climate” as a predictor without recording exposure does
not remedy missing information (Belykh & Vavilov, 2025; Fan et al., 2025; Kennedy & O’Hagan,
2001).
Sources of error and uncertainty quantification
Measurement error, parameter uncertainty, structural discrepancy and between-individual variation
should be distinguished. In Newtonian inversion, holding 𝑘, 𝑇0 and 𝑇𝑎 constant, sensitivity to 𝑇 is:
∂𝑡
∂𝑇 = − 1
𝑘(𝑇 − 𝑇𝑎). (9)
This derivation shows that near thermal equilibrium a small temperature perturbation may cause a large
change in inferred time. More thermometer decimal places do not remove this information limitation.
When several initial or environmental conditions are unknown, different histories may explain similar
observations; identifiability should therefore be examined before parameters are assigned unique
interpretations (Henssge & Madea, 2007; Sullivan, 2015; Wilk et al., 2022).
pág. 6021
A useful formulation is 𝑦𝑗 = ℋ[𝑇(𝐱, 𝑡𝑗; 𝜙)] + 𝛿𝑗 + 𝜖𝑗, where ℋ represents the measuring device and
location, 𝜙 the parameters, 𝛿𝑗 model discrepancy and 𝜖𝑗 observation error. This proposed inferential
organization draws on simulator calibration. Separating these terms prevents all disagreement from
being attributed to instrument noise or compensated for through physically implausible parameter
values (Kennedy & O’Hagan, 2001; Sullivan, 2015).
A forensic report should retain raw data, measurement locations and times, instrument calibration,
model version, environmental assumptions, point estimate and uncertainty range. An abstention rule is
proposed for cases outside the validated domain. Reporting should allow another professional to
reproduce the calculation and understand the assumptions supporting the interval. These
recommendations derive from predictive transparency; they are not presented as a universal legal
standard (Collins et al., 2015; Collins et al., 2024; Rudin, 2019).
Comparative synthesis and model map
Table 1 Conceptual comparison of models and scope of evidence
Model or
approach
Equations or
architecture
Required inputs Relevant
validation
Main limitations Metrics and
interpretation
Newton Lumped ODE,
exponential
Initial, body and
ambient
temperatures; 𝑘
Physical
idealization; no
universal cadaver
validation
Thermal uniformity,
constant
environment, initial
phase
Time error
conditional on
assumptions
Henssge Biexponential
relation and
empirical
corrections
Rectal and
ambient
temperatures,
mass, insulation
Experiments and
human application
within a restricted
domain
Environmental
history, initial
temperature,
corrections
Method-specific
intervals; resolution
is not accuracy
Numerical
PDEs
Distributed
conduction; finite
differences or
elements
Geometry,
properties,
boundaries, serial
temperatures
Controlled human
cadavers; specific
designs
Unknown
parameters, mesh,
anatomical
discrepancy
Both thermal
residuals and PMI
errors are needed
ML/DL Trees, regression,
networks and
ensembles
Metabolomics,
RNA, images,
context
Animal and human
evidence
distinguished by
study
Leakage, labels,
transportability
MAE/RMSE for
regression;
AUC/accuracy for
classes
PINN Network with
PDE residual and
auxiliary
conditions
Thermal data,
equation,
boundaries and
parameters
Computation and
engineering; no
direct forensic
validation in this
selection
Identifiability and
training; imposed
physics may be
wrong
Solution error and
residual are not
forensic error
Surrogate
hybrid
Physical
simulator and
cost-function
approximation
Serial
temperatures,
geometry,
parameter
bounds
Four human
cadavers in Wilk et
al.
Small sample;
dependence between
locations
Signed error 0.18 h;
SD 0.77 h; not MAE
Note. Synthesis based on Henssge, Otatsume, Wilk, Raissi, Cai and the AI reviews (Bugelli et al., 2026; Cai et al., 2021;
Henssge, 1988; Otatsume et al., 2024; Raissi et al., 2019; Sirago et al., 2026; Wilk et al., 2020; Wilk et al., 2022). ML: machine
learning; DL: deep learning; ODE: ordinary differential equation; PDE: partial differential equation; SD: standard deviation.
Metrics retain their original meanings.
pág. 6022
Figure 1 Proposed inference and validity-assessment workflow
Data and context acquisition, preprocessing by individual, physical, AI or hybrid models, temporal inference, uncertainty
assessment and conditional forensic use.
Note. Conceptual design based on thermal modeling and predictive reporting recommendations (Collins et al., 2024; Wilk et
al., 2022). The workflow is a methodological proposal, not an already validated clinical or forensic system. When evaluation
identifies insufficient data or a case outside the domain, the appropriate response is to widen the interval, request additional
information or abstain from an automated estimate.
The conceptual map separates two questions: what structure the model uses and where it has been tested.
“Physical,” “statistical” and “hybrid” describe construction; “animal,” “controlled human” and
“external forensic” describe evidence. Confusing these axes assigns an architecture a maturity level that
only validation can establish. Comparison must retain both axes alongside the temporal window and
the resolution required by the investigation (Karniadakis et al., 2021; Sirago et al., 2026; Wolff et al.,
2019).
CONCLUSIONS
Differential equations and artificial intelligence offer complementary approaches to PMI estimation:
the former make mechanisms and constraints explicit, while the latter relate high-dimensional signals
to temporal information.
pág. 6023
Their integration is scientifically defensible when learned structure, parameters and boundary
conditions are tested. The reviewed evidence does not establish general superiority of one family or
equivalence between experimental performance and forensic validity. Advances in cadaver simulation,
human metabolomics and surrogate-based optimization must be interpreted within their populations,
time windows and designs. Meaningful progress consists of reproducible, calibrated and transportable
inferences with uncertainty made sufficiently explicit to prevent an apparently precise prediction from
becoming an unjustified temporal claim.
Review limitations
The corpus was assembled through critical selection and targeted supplementation, not exhaustive
reproducible retrieval across every planned database. Individual verification of bibliographic identity
does not amount to full-text reading of every publication. Some extraction relied on available abstracts;
sample size, partitioning or methodological safeguards may therefore be described more fully in
inaccessible texts. The absence of independent review and a prospective protocol permits selection and
interpretation bias.
The 60 references include primary studies, reviews and methodological foundations; they are not 60
forensic trials. Heterogeneity prevents a single quantitative ranking or valid meta-analysis of temporal
errors. Regulations governing evidential admissibility were not exhaustively assessed across
jurisdictions. The absence of forensic PINN validation in this selection limits the corpus; it is not a
claim of universal nonexistence. Proposed tropical validation and algorithmic abstention require
empirical evaluation before recommendation as operational procedures.
Implications for future research
A staged research program should compare Henssge, distributed models, parsimonious statistical
predictors and hybrid systems on the same individuals and time windows. Before analysis, the design
should define decision-relevant errors, reference standards, partitioning units and procedures for
uncertain intervals. Sample-size planning must be based on independent individuals and effective
complexity; formulas developed for regression cannot be mechanically transferred to deep networks
(Collins et al., 2024; Riley et al., 2019).
pág. 6024
Prospective acquisition of thermal series and documented environments would better separate
parameters from discrepancy. External evaluation should include different institutions, instruments,
seasons and storage conditions. Hybrid models require ablation studies comparing data alone, physics
alone and their coupling with equal access to information. Improvements should persist outside the
dataset used to choose architecture, losses or hyperparameters (Karniadakis et al., 2021; S. Wang et al.,
2021; Wilk et al., 2022).
Research must also define when information cannot support a useful estimate. A method acknowledging
its application limits is preferable to a narrow interval without demonstrated coverage. The goal should
be to reduce relevant uncertainty and verify that reduction in new cases, rather than increase algorithmic
complexity or present as exact a chronology that the data do not identify (Gneiting & Raftery, 2007;
Kennedy & O’Hagan, 2001; Rudin, 2019; Sullivan, 2015).
REFERENCES
Althaus, L., & Henssge, C. (1999). Rectal temperature time of death nomogram: Sudden change of
ambient temperature. Forensic Science International, 99(3), 171–178.
https://doi.org/10.1016/s0379-0738(98)00188-1
An, G., Gao, Y., Cheng, S., Li, N., Ren, K., Du, Q., Bai, R., & Sun, J. (2025). Artificial intelligence in
forensic pathology: Multi-organ postmortem pathomics for estimating postmortem interval.
Computer Methods and Programs in Biomedicine, 270, 108949.
https://doi.org/10.1016/j.cmpb.2025.108949
Belykh, S. A., & Vavilov, A. Yu. (2025). Determination of «solar factor» magnitude in mathematical
simulation of the process of corpse’s temperature change of under conditions of insolation.
Sudebno-Meditsinskaya Ekspertiza, 68(2), 25–30.
https://doi.org/10.17116/sudmed20256802125
Bland, J. M., & Altman, D. G. (1986). Statistical methods for assessing agreement between two methods
of clinical measurement. The Lancet, 327(8476), 307–310. https://doi.org/10.1016/s0140-
6736(86)90837-8
pág. 6025
Bovenschen, M., Schwender, H., Ritz-Timme, S., Beseoglu, K., & Hartung, B. (2021). Estimation of
time since death after a post-mortem change in ambient temperature: Evaluation of a back-
calculation approach. Forensic Science International, 319, 110656.
https://doi.org/10.1016/j.forsciint.2020.110656
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
https://doi.org/10.1023/a:1010933404324
Brunton, S. L., Proctor, J. L., & Kutz, J. N. (2016). Discovering governing equations from data by
sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of
Sciences, 113(15), 3932–3937. https://doi.org/10.1073/pnas.1517384113
Bugelli, V., Calabrò, F., Camatti, J., Cecchi, R., Di Paolo, M., & Franceschetti, L. (2026). Artificial
intelligence in forensic science: A systematic review. Part II: Long-range postmortem interval
estimation. International Journal of Legal Medicine, 140(5), 3331–3343.
https://doi.org/10.1007/s00414-026-03856-4
Cai, S., Wang, Z., Wang, S., Perdikaris, P., & Karniadakis, G. E. (2021). Physics-informed neural
networks for heat transfer problems. Journal of Heat Transfer, 143(6), 060801.
https://doi.org/10.1115/1.4050542
Cianci, V., Mondello, C., Diaz, F. J., Zinger, T., Giese, K., Ryan, W., Češpivová, M., Sapienza, D.,
Gualniera, P., Asmundo, A., & Germanà, A. (2026). Temporal RT-qPCR-based porcine cardiac
molecular profiling for post-mortem interval estimation: Predictive modeling. International
Journal of Molecular Sciences, 27(11), 4856. https://doi.org/10.3390/ijms27114856
Cianci, V., Mondello, C., Sapienza, D., Guerrera, M. C., Cianci, A., Cracò, A., Omero, F., Gioffrè, V.,
Gualniera, P., Asmundo, A., & Germanà, A. (2024). Potential role of mRNA in estimating
postmortem interval: A systematic review. International Journal of Molecular Sciences,
25(15), 8185. https://doi.org/10.3390/ijms25158185
Collins, G. S., Moons, K. G. M., Dhiman, P., Riley, R. D., Beam, A. L., Van Calster, B., Ghassemi, M.,
Liu, X., Reitsma, J. B., van Smeden, M., Boulesteix, A.-L., Camaradou, J. C., Celi, L. A.,
Denaxas, S., Denniston, A. K., Glocker, B., Golub, R. M., Harvey, H., Heinze, G., Logullo, P.
pág. 6026
(2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that
use regression or machine learning methods. BMJ, 385, e078378.
https://doi.org/10.1136/bmj-2023-078378
Collins, G. S., Reitsma, J. B., Altman, D. G., & Moons, K. G. M. (2015). Transparent reporting of a
multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD
statement. Annals of Internal Medicine, 162(1), 55–63. https://doi.org/10.7326/m14-0697
Cuba-Gutierrez, L., Invernón-Monedero, M., Osuna, E., & Hernández-Romero, D. (2026). Application
of artificial intelligence in the determination of the postmortem interval: Systematic review of
the literature and metaanalysis. Forensic Science, Medicine and Pathology.
https://doi.org/10.1007/s12024-026-01309-3
Cui, C., Song, Y., Mao, D., Cao, Y., Qiu, B., Gui, P., Wang, H., Zhao, X., Huang, Z., Sun, L., & Zhong,
Z. (2023). Predicting the postmortem interval based on gravesoil microbiome data and a
random forest model. Microorganisms, 11(1), 56.
https://doi.org/10.3390/microorganisms11010056
Fan, W., Dai, X., Ye, Y., Yang, H., Sun, Y., Wu, J., Fu, Y., Shi, K., Chen, X., & Liao, L. (2025).
Estimation of postmortem interval under different ambient temperatures based on multi-organ
metabolomics and machine learning algorithm. International Journal of Legal Medicine,
139(5), 2561–2575. https://doi.org/10.1007/s00414-025-03523-0
Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal
of the American Statistical Association, 102(477), 359–378.
https://doi.org/10.1198/016214506000001437
Henssge, C. (1988). Death time estimation in case work. I. The rectal temperature time of death
nomogram. Forensic Science International, 38(3-4), 209–236. https://doi.org/10.1016/0379-
0738(88)90168-5
Henssge, C., & Madea, B. (2007). Estimation of the time since death. Forensic Science International,
165(2-3), 182–184. https://doi.org/10.1016/j.forsciint.2006.05.017
pág. 6027
Henssge, C., Madea, B., & Gallenkemper, E. (1988). Death time estimation in case work. II. Integration
of different methods. Forensic Science International, 39(1), 77–87.
https://doi.org/10.1016/0379-0738(88)90120-x
Ivanov, J., Ondruschka, B., & Heinemann, A. (2025). Einsatz eines 3D-Convolutional Neural Network
zur Prädiktion des postmortalen Intervalls in der postmortalen Computertomographie.
Rechtsmedizin, 35(4), 213–218. https://doi.org/10.1007/s00194-025-00765-5
Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-
informed machine learning. Nature Reviews Physics, 3(6), 422–440.
https://doi.org/10.1038/s42254-021-00314-5
Kennedy, M. C., & O’Hagan, A. (2001). Bayesian calibration of computer models. Journal of the Royal
Statistical Society Series B: Statistical Methodology, 63(3), 425–464.
https://doi.org/10.1111/1467-9868.00294
Klontzas, M. E., Leventis, D., Spanakis, K., Karantanas, A. H., & Kranioti, E. F. (2023). Post-mortem
CT radiomics for the prediction of time since death. European Radiology, 33(11), 8387–8395.
https://doi.org/10.1007/s00330-023-09746-2
Li, J., Wu, Y., Lu, X., Zhang, S., Li, N., Du, Q., Cao, J., Bai, R., & Sun, J. (2026). Semi-supervised
learning with dynamic classifier selection for PMI estimation: An animal study. Journal of
Forensic and Legal Medicine, 117, 103051. https://doi.org/10.1016/j.jflm.2025.103051
Liang, X., Deng, M., Zhu, Z., Zhang, W., Li, Y., Luo, J., Wang, H., Wu, S., Chen, R., Wang, G., Wu,
H., Shen, C., Hu, G., Zhang, K., Sun, Q., & Wang, Z. (2025). A novel approach for estimating
postmortem intervals under varying temperature conditions using pathology images and
artificial intelligence models. International Journal of Legal Medicine, 139(4), 1809–1819.
https://doi.org/10.1007/s00414-025-03447-9
Løber, I. M. M., Hedemann, M. S., Villesen, P., & Nielsen, K. L. (2025). Untangling the postmortem
metabolome: A machine learning approach for accurate PMI estimation. Analytical Chemistry,
97(30), 16123–16132. https://doi.org/10.1021/acs.analchem.4c05796
Lu, L., Meng, X., Mao, Z., & Karniadakis, G. E. (2021). DeepXDE: A deep learning library for solving
differential equations. SIAM Review, 63(1), 208–228. https://doi.org/10.1137/19m1274067
pág. 6028
Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J.,
Bansal, N., & Lee, S.-I. (2020). From local explanations to global understanding with
explainable AI for trees. Nature Machine Intelligence, 2(1), 56–67.
https://doi.org/10.1038/s42256-019-0138-9
Magnusson, R., Söderberg, C., Ward, L. J., Arpe, J., Kugelberg, F. C., Elmsjö, A., Green, H., & Nyman,
E. (2026). The human metabolome and machine learning improves predictions of the post-
mortem interval. Nature Communications, 17(1), 1504. https://doi.org/10.1038/s41467-026-
69158-w
Metcalf, J. L., Xu, Z. Z., Weiss, S., Lax, S., Van Treuren, W., Hyde, E. R., Song, S. J., Amir, A., Larsen,
P., Sangwan, N., Haarmann, D., Humphrey, G. C., Ackermann, G., Thompson, L. R., Lauber,
C., Bibat, A., Nicholas, C., Gebert, M. J., Petrosino, J. F., … Knight, R. (2016). Microbial
community assembly and metabolic function during mammalian corpse decomposition.
Science, 351(6269), 158–162. https://doi.org/10.1126/science.aad2646
Nedugov, G. V. (2021a). Double exponential model of corpse cooling under conditions of linearly
varying ambient temperature. Russian Journal of Forensic Medicine, 7(4), 19–28.
https://doi.org/10.17816/fm429
Nedugov, G. V. (2021b). New computer technologies to determine postmortem interval by the Henssge
method. Russian Journal of Forensic Medicine, 7(3), 152–158. https://doi.org/10.17816/fm406
Nedugov, V. G., & Nedugov, G. V. (2024). Thermometrical determination of the postmortem interval
by non-linear optimization method. Sudebno-Meditsinskaya Ekspertiza, 67(5), 19–23.
https://doi.org/10.17116/sudmed20246705119
Nedugov, V. G., & Nedugov, G. V. (2025). Mathematical programming in assessment of the
postmortem interval under conditions of linearly varying external temperature. Sudebno-
Meditsinskaya Ekspertiza, 68(4), 34–39. https://doi.org/10.17116/sudmed20256804134
Otatsume, M., Shinkawa, N., Tachibana, M., Kuroki, H., Ro, A., Sonoda, A., Kakizaki, E., & Yukawa,
N. (2024). Technical note: Excel spreadsheet calculation of the Henssge equation as an aid to
estimating postmortem interval. Journal of Forensic and Legal Medicine, 101, 102634.
https://doi.org/10.1016/j.jflm.2023.102634
pág. 6029
Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., Shamseer,
L., Tetzlaff, J. M., Akl, E. A., Brennan, S. E., Chou, R., Glanville, J., Grimshaw, J. M.,
Hróbjartsson, A., Lalu, M. M., Li, T., Loder, E. W., Mayo-Wilson, E., McDonald, S., … Moher,
D. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic
reviews. BMJ, 372, n71. https://doi.org/10.1136/bmj.n71
Pennes, H. H. (1948). Analysis of tissue and arterial blood temperatures in the resting human forearm.
Journal of Applied Physiology, 1(2), 93–122. https://doi.org/10.1152/jappl.1948.1.2.93
Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep
learning framework for solving forward and inverse problems involving nonlinear partial
differential equations. Journal of Computational Physics, 378, 686–707.
https://doi.org/10.1016/j.jcp.2018.10.045
Ren, L., Wang, Z., Yang, Z., Tian, R., Du, J., Dang, Z., Li, J., Guo, X., Xia, Y., Cui, W., & Wang, Y.
(2026). Microbial succession and machine learning for postmortem interval estimation in
bromadiolone-poisoned cadavers. AMB Express, 16(1), 26. https://doi.org/10.1186/s13568-
026-02023-7
Riley, R. D., Snell, K. I. E., Ensor, J., Burke, D. L., Harrell, F. E., Moons, K. G. M., & Collins, G. S.
(2019). Minimum sample size for developing a multivariable prediction model: Part I –
Continuous outcomes. Statistics in Medicine, 38(7), 1262–1275.
https://doi.org/10.1002/sim.7993
Roberts, M., Driggs, D., Thorpe, M., Gilbey, J., Yeung, M., Ursprung, S., Aviles-Rivero, A. I., Etmann,
C., McCague, C., Beer, L., Weir-McCall, J. R., Teng, Z., Gkrania-Klotsas, E., AIX-COVNET,
Ruggiero, A., Korhonen, A., Jefferson, E., Ako, E., Langs, G., … Schönlieb, C.-B. (2021).
Common pitfalls and recommendations for using machine learning to detect and prognosticate
for COVID-19 using chest radiographs and CT scans. Nature Machine Intelligence, 3(3), 199–
217. https://doi.org/10.1038/s42256-021-00307-0
Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use
interpretable models instead. Nature Machine Intelligence, 1(5), 206–215.
https://doi.org/10.1038/s42256-019-0048-x
pág. 6030
Sharma, R., Diksha, Bhute, A. R., & Bastia, B. K. (2022). Application of artificial intelligence and
machine learning technology for the prediction of postmortem interval: A systematic review of
preclinical and clinical studies. Forensic Science International, 340, 111473.
https://doi.org/10.1016/j.forsciint.2022.111473
Shen, Z., Zhong, Y., Wang, Y., Zhu, H., Liu, R., Yu, S., Zhang, H., Wang, M., Yang, T., & Zhang, M.
(2024). A computational approach to estimate postmortem interval using postmortem computed
tomography of multiple tissues based on animal experiments. International Journal of Legal
Medicine, 138(3), 1093–1107. https://doi.org/10.1007/s00414-023-03127-6
Sirago, G., Solarino, B., Dell’Erba, A., & Ferorelli, D. (2026). Artificial intelligence for short-term
post-mortem interval estimation: A systematic review with PROBAST and TRIPOD+AI
appraisal. Legal Medicine, 85, 102931. https://doi.org/10.1016/j.legalmed.2026.102931
Stigter, H., Krap, T., & Duijst, W. L. J. M. (2024). Estimation of the post-mortem interval; added value
of mechanical excitability of human skeletal muscle. Journal of Forensic and Legal Medicine,
103, 102664. https://doi.org/10.1016/j.jflm.2024.102664
Sullivan, T. J. (2015). Introduction to uncertainty quantification. Springer. https://doi.org/10.1007/978-
3-319-23395-6
Teixeira, M. J., Barbosa, D. J., Dinis-Oliveira, R. J., & Freitas, A. R. (2025). Redefining postmortem
interval estimation: The need for evidence-based research to bridge science and justice.
Frontiers in Microbiology, 16, 1646907. https://doi.org/10.3389/fmicb.2025.1646907
Tournois, L., Trousset, V., Hatsch, D., Delabarde, T., Ludes, B., & Lefèvre, T. (2024). Artificial
intelligence in the practice of forensic medicine: A scoping review. International Journal of
Legal Medicine, 138(3), 1023–1037. https://doi.org/10.1007/s00414-023-03140-9
Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L., Steyerberg, E. W., & Topic Group
‘Evaluating diagnostic tests and prediction models’ of the STRATOS initiative. (2019).
Calibration: The Achilles heel of predictive analytics. BMC Medicine, 17(1), 230.
https://doi.org/10.1186/s12916-019-1466-7
pág. 6031
Wang, S., Teng, Y., & Perdikaris, P. (2021). Understanding and mitigating gradient flow pathologies
in physics-informed neural networks. SIAM Journal on Scientific Computing, 43(5), A3055–
A3081. https://doi.org/10.1137/20m1318043
Wang, Y., Li, B., Wang, Y., Chen, Q., Wang, Z., Fu, G., Li, S., Zhang, C., Zhou, Z., & Cong, B. (2026).
Endogenous and exogenous small RNA signatures as novel tools for postmortem interval
determination. Biomolecules, 16(3), 474. https://doi.org/10.3390/biom16030474
Wenzlow, N., Mills, D., Byrd, J., Warren, M., & Long, M. T. (2023). Review of the current and potential
use of biological and molecular methods for the estimation of the postmortem interval in
animals and humans. Journal of Veterinary Diagnostic Investigation, 35(2), 97–108.
https://doi.org/10.1177/10406387231153930
Wilk, L. S., Edelman, G. J., & Aalders, M. C. G. (2022). Next-generation time of death estimation:
Combining surrogate model-based parameter optimization and numerical thermodynamics.
Royal Society Open Science, 9(7), 220162. https://doi.org/10.1098/rsos.220162
Wilk, L. S., Hoveling, R. J. M., Edelman, G. J., Hardy, H. J. J., van Schouwen, S., van Venrooij, H., &
Aalders, M. C. G. (2020). Reconstructing the time since death using noninvasive thermometry
and numerical analysis. Science Advances, 6(22), eaba4243.
https://doi.org/10.1126/sciadv.aba4243
Wolff, R. F., Moons, K. G. M., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., Reitsma, J.
B., Kleijnen, J., & Mallett, S. (2019). PROBAST: A tool to assess the risk of bias and
applicability of prediction model studies. Annals of Internal Medicine, 170(1), 51–58.
https://doi.org/10.7326/m18-1376
Yang, L., Meng, X., & Karniadakis, G. E. (2021). B-PINNs: Bayesian physics-informed neural
networks for forward and inverse PDE problems with noisy data. Journal of Computational
Physics, 425, 109913. https://doi.org/10.1016/j.jcp.2020.109913
Yang, Y., & Perdikaris, P. (2019). Adversarial uncertainty quantification in physics-informed neural
networks. Journal of Computational Physics, 394, 136–152.
https://doi.org/10.1016/j.jcp.2019.05.027
pág. 6032
Yuan, H., Wang, Z., Wang, Z., Zhang, F., Guan, D., & Zhao, R. (2023). Trends in forensic
microbiology: From classical methods to deep learning. Frontiers in Microbiology, 14,
1163741. https://doi.org/10.3389/fmicb.2023.1163741