Detecting and Subtyping Ketoacidosis from Metabolomic Patterns in Forensic Casework

This study demonstrates that integrating postmortem metabolomics with machine learning on real Swedish forensic cases enables accurate detection and subtyping of ketoacidosis-related deaths, achieving over 90% detection accuracy and over 80% subtyping accuracy.

The Gist

Imagine a detective arriving at a crime scene. In the world of forensics, the "crime" is often a mysterious death where the body has shut down, but the reason isn't immediately obvious. One tricky culprit is ketoacidosis.

Think of ketoacidosis as a car engine that has run out of its preferred fuel (sugar/carbs) and has started burning its own body fat for energy. This process creates "exhaust fumes" called ketones. If these fumes build up too much, they turn the body's blood into acid, which can be fatal.

The problem? This engine trouble can happen for three very different reasons:

1. Diabetes: The body can't use sugar properly (Diabetic Ketoacidosis).
2. Alcoholism: The body is flooded with alcohol and lacks sugar (Alcoholic Ketoacidosis).
3. Hypothermia: The body is freezing and burning fat to stay warm.

Traditionally, forensic pathologists have to play a guessing game. They look for specific clues (like measuring sugar or ketones), but it's like trying to figure out if a car broke down because of a flat tire, a bad battery, or running out of gas, just by looking at the smoke. It's hard to tell the difference.

The New Detective Tool: The "Metabolic Fingerprint"

This paper introduces a high-tech solution: Machine Learning combined with Metabolomics.

Imagine that when a person dies, their body leaves behind a massive, complex "fingerprint" made of thousands of tiny chemical signals (metabolites) in their blood. It's like a chaotic library of books where every book represents a different chemical. A human detective can only read a few pages at a time. But a Machine Learning computer is like a super-fast librarian who can read the entire library in a split second, spot patterns, and tell you exactly which story the body was telling.

How They Did It

The researchers in Sweden gathered a massive library of these chemical fingerprints from 1,788 real forensic cases. They had:

• The "Sick" Group: People who died from the three types of ketoacidosis (Diabetes, Alcohol, or Hypothermia).
• The "Control" Group: People who died from hanging. They chose this group because hanging is usually a quick process, meaning the body didn't have time to change its chemical makeup much. They are the "baseline" or the "clean slate."

They fed this data into three different types of computer brains (algorithms):

1. Random Forest: Like a committee of experts voting on the answer.
2. LASSO: A strict editor that cuts out unnecessary information to find the most important clues.
3. SVM: A mathematical line-drawer that separates the "sick" from the "healthy" with perfect precision.

The Results: A Super-Smart Detective

The results were impressive. The computer models became expert detectives:

• Spotting the Problem: They could tell if a death was caused by ketoacidosis (any kind) versus a normal death with over 90% accuracy.
• Solving the Mystery: Even better, they could tell which type of ketoacidosis it was (Diabetes vs. Alcohol vs. Cold) with over 80% accuracy.
• The "Starvation" Test: To prove the models weren't just memorizing the answers, they tested them on a group of people who died from starvation (a condition they hadn't seen before). The models correctly identified that starvation also causes ketoacidosis, proving they understood the pattern, not just the specific cases.

The "Aha!" Moments

The computer didn't just guess; it told the scientists why it was guessing. By looking at the most important chemical clues, they found:

• Cortisol: A stress hormone that was high in all the ketoacidosis cases.
• Glucosamine: A chemical linked to diabetes, which helped the computer spot the diabetic cases.
• Pyridone: A chemical linked to Vitamin B3 breakdown, which helped spot the hypothermia (freezing) cases.

Why This Matters

Think of this as giving forensic pathologists a crystal ball. Instead of relying on a single test that might be ambiguous, they can now look at the entire chemical landscape of the body.

This is a game-changer because:

1. It's Objective: It removes human bias.
2. It's Fast: It can process data much quicker than a human can analyze a list of chemicals.
3. It's Comprehensive: It sees the whole picture, not just one or two symptoms.

The Bottom Line

This study shows that by teaching computers to read the body's chemical "language," we can solve cold cases of death much more accurately. It's like upgrading from a magnifying glass to a super-microscope that can see the invisible story of how and why a person's body shut down. While there is still work to be done (like translating these findings to other countries' labs), this is a giant leap forward for forensic science.

Technical Summary

1. Problem Statement

Determining the cause of death in forensic casework is particularly challenging when different conditions share similar pathophysiological mechanisms. Ketoacidosis, a metabolic state characterized by blood acidification due to ketone body accumulation, can arise from multiple distinct causes:

• Diabetic Ketoacidosis (DKA)
• Alcoholic Ketoacidosis (AKA)
• Hypothermia
• Starvation

Current forensic practices rely on targeted measurements of specific biomarkers (e.g., ketone bodies, glucose) in vitreous humor. However, these targeted approaches often fail to capture the broader metabolic context required to differentiate between these closely related subtypes or to distinguish ketoacidosis from other causes of death. There is a critical need for an objective, high-throughput approach that utilizes the full metabolic profile to improve cause-of-death determination.

2. Methodology

Data Source and Cohorts

The study utilized a unique, real-world dataset from the National Board of Forensic Medicine in Sweden, comprising postmortem femoral blood samples collected between 2017 and 2020.

• Total Samples: 1,788 cases.
• Development Cohort (n=1,698):
  • Ketoacidosis Cases (n=469): AKA (n=109), DKA (n=220), and Hypothermia (n=140).
  • Controls (n=1,229): Hanging cases (selected due to rapid death and minimal physiological alteration).
• Independent Validation Cohorts (n=90):
  • Starvation cases (n=21).
  • Alcoholic Controls (AC) (n=29): Alcohol as a secondary cause, main cause non-ketoacidosis.
  • Diabetic Controls (DC) (n=40): Diabetes as a secondary cause, main cause non-ketoacidosis.

Analytical Workflow

1. Data Acquisition: Samples underwent toxicological screening using UHPLC-QTOF (Ultra-High-Performance Liquid Chromatography coupled with Quadrupole Time-of-Flight Mass Spectrometry) in positive ionization mode.
2. Preprocessing:
   • Raw data processed using xcms and CAMERA packages in R.
   • Feature detection via the centWave algorithm.
   • Filtering criteria: Features present in at least 60% of cases in at least one group.
   • Result: 4,484 metabolic features extracted.
3. Statistical Analysis:
   • Differential Abundance: Mann-Whitney U tests with Bonferroni correction, adjusted for covariates (Postmortem Interval, BMI, Age, Gender) using linear modeling (limma).
   • Feature Selection: Features with $q < 0.05$ and $|Log2 FC| > 1.5$ were considered significant.
   • Annotation: Metabolites were annotated using MetaboAnalystR, KEGG, BiGG, and HMDB databases (201/4,484 features annotated).
4. Machine Learning Models:
   Three supervised learning algorithms were trained and compared for both Binary Classification (Ketoacidosis vs. Control) and Multinomial Classification (AKA vs. DKA vs. Hypothermia vs. Control):
   • Random Forest (RF)
   • LASSO (Least Absolute Shrinkage and Selection Operator)
   • Support Vector Machine (SVM) with linear kernel.
   • Training Strategy: 70% of the development cohort for training, 30% for testing. Independent cohorts were used strictly for validation without retraining.

3. Key Results

Metabolomic Signatures

• Differential Features: 1,416 features showed significant differences between ketoacidosis and controls. After covariate adjustment, 133 features were significantly higher and 37 were lower in ketoacidosis cases.
• PCA Limitations: Principal Component Analysis (PCA) failed to show distinct separation between groups (PC1: 7.4%, PC2: 4.1% variance), confirming the necessity of supervised ML for pattern recognition.

Binary Classification Performance (Ketoacidosis vs. Control)

• All three models (RF, LASSO, SVM) achieved high performance.
• Accuracy: >90% balanced accuracy.
• True Positive Rate (TPR): 80.9% – 89.4%.
• True Negative Rate (TNR): ~98%.
• Generalization: When applied to the independent Starvation cohort (n=21), the models correctly classified 81% (RF) to 90% (SVM) of starvation cases as ketoacidosis-related, demonstrating robust generalization to unseen metabolic states.

Multinomial Classification Performance (Subtyping)

• The models successfully distinguished between AKA, DKA, Hypothermia, and Controls.
• Balanced Accuracy: 0.83 – 0.88.
• Recall:
  • DKA: High recall (~84.8% in RF).
  • AKA: Moderate recall (~57.6% in RF).
  • Hypothermia: Struggled most, often misclassified as controls.
• Validation on Controls:
  • Alcoholic Controls (AC): ~70% correctly classified as controls (not AKA) by RF.
  • Diabetic Controls (DC): Mostly classified as controls, though some were misclassified as DKA, suggesting the model captures underlying diabetic pathology even when not the primary cause of death.

Feature Importance

Clustering of the top 25 features revealed distinct grouping patterns:

• Controls clustered separately from AKA/DKA.
• Hypothermia clustered closer to controls than to other ketoacidosis types.
• Key Annotated Metabolites:
  • Cortisol: Elevated in all ketoacidosis subtypes; known marker for hypothermia/stress.
  • Glucosamine: Higher in DKA and DC groups (linked to insulin resistance).
  • 2PY/4PY (Pyridone derivatives): Breakdown products of Vitamin B3/NAD; elevated in hypothermia.
  • 5,6-indolequinone-2-carboxylic acid: Elevated in hypothermia.

4. Key Contributions

1. First Real-World Application: This is the first study to integrate postmortem metabolomics with machine learning to detect and subtype ketoacidosis using real forensic casework rather than controlled clinical datasets.
2. Methodological Framework: Demonstrates that high-dimensional, untargeted metabolomics data from routine toxicological screening (UHPLC-QTOF) can be repurposed for cause-of-death determination.
3. Robust Validation: The use of independent validation cohorts (starvation, AC, DC) proves that the models generalize beyond the training data and can distinguish between primary causes of death and secondary comorbidities.
4. Subtyping Capability: Successfully differentiates between AKA, DKA, and hypothermia, which is currently difficult with standard targeted biomarkers.

5. Significance and Limitations

Significance:

• Forensic Utility: Provides an objective, data-driven tool to assist forensic pathologists in determining the cause of death in complex cases where traditional biomarkers are ambiguous.
• Scalability: Leverages existing routine forensic data (toxicology screens), making the approach cost-effective and scalable without requiring new sample collection.
• Mechanistic Insight: Identifies specific metabolites (e.g., cortisol, glucosamine) that serve as potential biomarkers for specific subtypes of ketoacidosis.

Limitations:

• Annotation Rate: Only ~4.5% (201/4,484) of features were successfully annotated, limiting deep mechanistic interpretation of the remaining "unknown" features.
• Ground Truth Reliability: Labels are based on the forensic pathologist's best determination, which may not always reflect the absolute biological truth.
• Generalizability: The dataset is specific to the Swedish forensic system and analytical pipeline; external validation on different instruments or jurisdictions is needed.
• Model Complexity: The study used linear kernels and standard tree-based models; more complex non-linear models (e.g., deep learning) were not explored due to data size constraints.

Conclusion:
The study concludes that combining supervised machine learning with postmortem metabolomics is a powerful, accurate, and robust method for detecting and subtyping ketoacidosis-related deaths, offering a significant advancement for forensic pathology.