Predictive Analytics for Foot Ulcers Using Time-Series Temperature and Pressure Data

This paper presents a predictive analytics framework that utilizes time-series temperature and pressure data from wearable foot sensors, processed through Isolation Forest and K-Nearest Neighbors algorithms, to detect early anomalies indicative of diabetic foot ulcer risk for real-time health surveillance.

The Gist

Imagine your feet are like the tires on a car. If you drive with a flat tire or if the engine gets too hot, you know something is wrong before the car breaks down completely. For people with diabetes, their feet are like those tires, but they often can't "feel" the warning signs because the nerves are damaged. A small blister or a hot spot can turn into a serious, life-threatening wound (an ulcer) before they even realize it.

This paper is about building a digital "check engine light" for diabetic feet.

Here is the story of how the researchers tried to build it, explained simply:

1. The Problem: The Silent Danger

Diabetes can cause two big problems in the feet:

• Numbness: You can't feel pain, so you don't notice a rock in your shoe or a blister forming.
• Circulation issues: Your feet might get cold or swollen, which hurts healing.

Usually, doctors only check feet once a year. But ulcers can start growing in the time between visits. The researchers wanted a way to watch the feet 24/7, like a security camera, to catch trouble before it becomes a disaster.

2. The Tools: Smart Socks and Sensors

The team didn't use X-rays or MRI machines. Instead, they used wearable sensors (think of them as high-tech insoles or smart socks).

• Temperature Sensors: Like a fever thermometer for your foot. If one spot gets hotter than the rest, it might mean inflammation (the body fighting a hidden injury).
• Pressure Sensors: Like a pressure pad. They measure how hard you are stepping on specific parts of your foot. If you are standing on one spot too hard for too long, the skin there could break down.

They collected data from healthy people walking on a special path to see what "normal" walking looks like.

3. The Brain: Two Different Detectives

The researchers used two different types of computer "detectives" (Machine Learning algorithms) to look at the data and find anything weird. They didn't teach the computers what an ulcer looks like (because they didn't have enough data of actual ulcers). Instead, they taught them what normal looks like and asked them to shout, "Hey, that doesn't look right!"

Detective A: The "Isolation Forest" (The Subtle Observer)

• How it works: Imagine a forest of trees. This detective isolates data points by chopping the forest into smaller and smaller pieces. If a piece of data is weird, it gets isolated quickly.
• Its Superpower: It is great at spotting tiny, subtle changes. It's like a detective who notices you are walking slightly differently or that your foot is 0.5 degrees warmer than usual.
• The Result: This detective was very good at finding the early warning signs of trouble. It didn't scream too often, but when it did, it was usually right about something small starting to go wrong.

Detective B: The "K-Nearest Neighbors" (The Alarmist)

• How it works: This detective looks at a data point and asks, "Who are your 20 closest friends?" If your friends are all normal, but you are way out of line, you are an outlier.
• Its Superpower: It is great at spotting huge, dramatic changes. It's like a smoke alarm that goes off when there is a massive fire.
• The Result: This detective was very sensitive to extreme events (like a sudden, massive spike in heat or pressure). However, it was a bit of a "cry wolf." It sometimes flagged things as dangerous when they were just a little weird, which could annoy patients with too many false alarms.

4. The Big Discovery: The "Hot and Heavy" Connection

The researchers found something interesting: Pressure and Temperature are best friends.
When the foot is under too much pressure (heavy stepping), that spot tends to get hotter. It's like rubbing your hands together quickly; they get hot from the friction.

• The Lesson: You can't just watch the heat or just watch the pressure. You need to watch both at the same time. If a spot is getting hot and heavy, that's a double warning sign that an ulcer might be forming.

5. The Verdict: What Works Best?

• The Winner for Early Warning: The Isolation Forest detective. It's the best at catching the problem when it's just a tiny spark, allowing doctors to intervene before a fire starts.
• The Role of the Other: The KNN detective is good for spotting emergencies, but it might be too noisy for everyday use.

6. The Catch (Limitations)

There is one big "but." The researchers tested this on healthy people in a lab.

• The Reality Check: Real diabetic feet are different. They might be colder, have different circulation, or the sensors might slip. Also, the "smart socks" need to be comfortable enough that people actually want to wear them every day.
• The Future: The next step is to test this on real patients with diabetes, in their real homes, to see if the "check engine light" works when the car is actually breaking down.

Summary

This paper is about using smart sensors and computer detectives to watch diabetic feet 24/7. By combining temperature and pressure data, they found a way to spot the tiny, early signs of a foot ulcer before it becomes a serious wound. It's like having a guardian angel that whispers, "Hey, your foot is getting a little hot and heavy, let's change your shoes," before the damage is done.

Technical Summary

1. Problem Statement

Diabetic Foot Ulcers (DFUs) are a severe complication of diabetes, leading to high rates of amputation, mortality, and significant healthcare costs (approx. £837–£962 million annually in the UK). Current monitoring methods rely on sporadic clinical evaluations and subjective self-reports, which fail to provide the continuous, real-time data necessary for early intervention.

• The Gap: Existing risk models often depend on supervised learning requiring large, labeled datasets of ulceration events, which are difficult to collect. Furthermore, many commercial sensors neglect plantar shear stress and fail to capture subtle, pre-clinical physiological changes (e.g., micro-inflammation or minor pressure shifts) before tissue breakdown occurs.
• Objective: To develop a predictive analytics framework using unsupervised machine learning on continuous time-series data from wearable sensors to detect anomalies indicative of early ulcer risk without requiring pre-labeled ulcer data.

2. Methodology

Data Collection and Preprocessing

• Data Source: Time-series data was collected from healthy adults walking on an instrumented pathway.
• Sensors:
  • Temperature: 10,000 NTC thin-film thermocouples.
  • Pressure: FlexiForce A401 pressure sensors for plantar load monitoring.
• Preprocessing:
  • Handling Missing Data: Utilized Complete Case Analysis, Forward-Filling, and Zero-Filling, with a "filled flag" to track imputation.
  • Outlier Removal: Applied the Interquartile Range (IQR) method (upper quartile set at the 85th percentile) to filter noise while preserving gait variability.
  • Activity Windowing: Data was truncated to include only "foot-ground contact" periods, identified by a pressure threshold at the 95th percentile.

Feature Engineering

Features were extracted to capture physiological shifts:

• Pressure Features: Maximum/mean pressure, step count, step period, and Area Under the Curve (AUC) for pressure.
• Temperature Features: Maximum/minimum/mean temperatures, temperature change rates, and peak counts.
• Selection: Redundant features (variance < 0.01) and highly correlated features (correlation > 0.95) were removed to address multicollinearity.

Anomaly Detection Algorithms

The study employed two unsupervised algorithms to identify deviations from normal gait patterns:

1. Isolation Forest (ISO): A partitioning-based ensemble method.
   • Configuration: 100 trees, max samples 0.6, contamination rate 0.05.
   • Mechanism: Detects anomalies by isolating observations via recursive splitting; lower anomaly scores indicate higher likelihood of risk.
2. K-Nearest Neighbors (KNN) / Local Outlier Factor (LOF): A distance/density-based method.
   • Configuration: $k=20$ neighbors, contamination rate 0.05.
   • Mechanism: Calculates the local density deviation of a point relative to its neighbors; effective at flagging extreme deviations.

3. Key Contributions

• Unsupervised Framework: Demonstrated the viability of using unsupervised learning for DFU prediction, bypassing the need for scarce labeled datasets of actual ulcers.
• Multi-Modal Sensor Fusion: Validated the correlation between plantar pressure and foot temperature, arguing for a combined monitoring approach rather than single-modality analysis.
• Algorithmic Comparison: Provided a rigorous comparative analysis of Isolation Forest vs. KNN in the context of biomechanical time-series data, highlighting their distinct strengths and weaknesses.
• Feature Importance Mapping: Identified specific sensors (e.g., Sensor 3 for midfoot/heel) and features (e.g., step count, AUC) that are most predictive of ulcer risk.

4. Results Analysis

Performance Comparison

• Isolation Forest:
  • Strength: Highly sensitive to micro-anomalies and subtle deviations (e.g., slight inflammation, minor prolonged stress).
  • Utility: Ideal for early-stage detection and proactive intervention before clinical symptoms appear.
  • False Positives: Lower rate compared to KNN.
• K-Nearest Neighbors (LOF):
  • Strength: Highly sensitive to extreme deviations (e.g., sudden temperature spikes, massive pressure surges).
  • Weakness: Higher false-positive rate, which could lead to "alert fatigue" in a real-time clinical setting.
  • Utility: Better suited as a complementary tool for flagging severe, acute events rather than continuous monitoring.

Key Findings

• Correlation: A positive correlation was found between pressure and temperature (e.g., max pressure on Sensor 3 correlated with max temperature at 0.41–0.48), suggesting that increased mechanical stress leads to localized heating (inflammation/friction).
• Temporal Patterns: Anomalies were concentrated in specific periods (Sept–Oct 2023, Jan 2024, April 2024).
  • Seasonal Impact: A significant temperature drop in December 2023 suggested vasoconstriction due to cold weather, a known risk factor for DFUs.
  • Pressure Peaks: Maximum pressure peaked in Jan and June 2024, while extended step periods (indicating sustained load) were noted in Nov 2023 and Feb 2024.
• Sensor Specificity: Sensor 3 (likely midfoot/heel) and Sensor 4 were critical for pressure anomaly detection, while maximum temperature spikes were the primary indicators for thermal anomalies.

5. Significance and Limitations

Significance

• Clinical Impact: The framework enables real-time surveillance, allowing clinicians to intervene before tissue breakdown occurs, potentially reducing amputation rates and healthcare costs.
• Scalability: The unsupervised approach is scalable to large populations without the bottleneck of manual labeling.
• Preventive Care: Shifts the paradigm from reactive treatment to proactive prevention by identifying subtle physiological precursors to ulcers.

Limitations

• Dataset Scope: Data was collected from healthy adults in a controlled lab environment. The models have not yet been validated on diabetic patients with varying stages of neuropathy or comorbidities.
• Environmental Factors: Variables such as footwear type, ambient humidity, and daily activity variations were not integrated, which could affect sensor readings in real-world scenarios.
• Algorithmic Noise: The KNN model's high false-positive rate requires refinement (e.g., ensemble methods) before clinical deployment to avoid alert fatigue.
• Generalizability: The transition from a controlled instrumented pathway to unstructured daily walking requires further validation regarding sensor placement consistency and patient compliance.

Conclusion

The paper establishes a foundational framework for using wearable sensor data and unsupervised machine learning to predict Diabetic Foot Ulcers. While Isolation Forest emerges as the superior candidate for early, continuous monitoring due to its sensitivity to subtle changes, the study advocates for a multi-modal approach combining pressure and temperature data. Future work must focus on validating these models with diverse diabetic populations and integrating environmental context to ensure clinical viability.