ML Trauma Assessment

1. Project Title

Development of an Artificial Intelligence Machine Learning EMS Injury Assessment Prototype Combining Computer Vision, Machine Learning, and Protocol-Based Information

2. Project Overview

The AI EMS Assessment system was developed as an experimental software prototype designed to demonstrate how artificial intelligence could combine image classification and patient information within a single EMS-oriented interface.

The system combines three primary components:

• Computer vision for classification of injury photographs.
• Machine learning for a prototype risk estimate based on patient information and vital signs.
• Protocol-based information associated with the predicted injury category.

The system was implemented using Python and Streamlit. The computer-vision component uses a ResNet-18 neural network architecture through the timm library. The risk component uses a previously trained machine-learning model loaded through joblib.

The prototype was developed for educational and research purposes and was not designed or validated as an autonomous clinical decision-making system.

3. Development Objective

The primary development objective was to create a functional prototype capable of accepting two forms of information:

Patient information

• Age
• Pain score
• Heart rate
• Systolic blood pressure
• Diastolic blood pressure
• SpO₂
• Respiratory rate
• GCS

Visual information

• Injury photograph

The system processes these inputs and produces an integrated output containing:

• Predicted injury classification
• Model confidence
• Classification probabilities
• Patient vital signs
• Prototype risk estimate
• Screening flags
• Protocol-based considerations
• Model limitations

4. System Architecture

The prototype follows the general architecture:

USER INPUT

│

┌────────────┴────────────┐

│ │

Patient Data Injury Image

│ │

│ ▼

│ Image Preprocessing

│ │

│ ▼

│ ResNet-18 Model

│ │

│ ▼

│ Injury Classification

│ │

▼ │

Risk Model │

│ │

▼ │

Risk Estimate │

│ │

└────────────┬────────────┘

│

▼

Integrated Assessment

│

┌─────────┼─────────┐

│ │ │

▼ ▼ ▼

Findings Protocol Confidence

The design intentionally separates the computer-vision model from the risk model. Each model performs a different computational task.

5. Software Development Environment

The prototype was implemented in Python.

The primary libraries used were:

Library	Function
Streamlit	Web-based user interface
PyTorch	Neural-network inference
torchvision	Image preprocessing
timm	ResNet-18 model architecture
pandas	Patient-data organization
Pillow	Image loading and conversion
joblib	Loading the risk model
protocols.py	Protocol-information system

6. User Interface Development

Streamlit was selected to provide a simple interface between the user and the underlying machine-learning models.

The application was configured with:

• Wide page layout
• EMS-themed title
• Sidebar for patient information
• Injury-image upload area
• Assessment button
• Results sections
• Warning messages
• Model limitation information

The application begins by displaying an educational/research warning indicating that the system has not been clinically validated.

This warning is an important part of the prototype design because the computational outputs are not intended to be represented as established medical diagnoses.

7. Patient Input Module

The patient-input module was developed using Streamlit input controls.

The user can enter:

Age

Pain

Heart Rate

Systolic Blood Pressure

Diastolic Blood Pressure

SpO₂

Respiratory Rate

GCS

Each variable has an appropriate numerical range within the interface.

For example, heart rate is restricted to a range of 0–300, while GCS is restricted to 3–15.

The values are collected and organized into a pandas DataFrame.

The resulting structure is:

age

heart_rate

systolic_bp

diastolic_bp

spo2

respiratory_rate

gcs

pain

This structure is passed to the trained risk model.

8. Injury Image Module

The image-analysis component allows the user to upload an injury photograph.

The application accepts:

.jpg

.jpeg

.png

When an image is uploaded, it is opened using Pillow and converted to RGB format.

This standardizes the image representation before it is passed to the neural network.

9. Image Preprocessing

Before being analyzed by the neural network, each image passes through a preprocessing pipeline.

The image is resized to:

224 × 224 pixels

It is then converted to a PyTorch tensor.

Finally, the tensor is normalized using:

Mean:

[0.485, 0.456, 0.406]

Standard deviation:

[0.229, 0.224, 0.225]

The resulting tensor is expanded to include a batch dimension before being passed to the neural network.

The preprocessing pipeline can be represented as:

Original Image

↓

RGB Conversion

↓

224 × 224 Resize

↓

Tensor Conversion

↓

Normalization

↓

Batch Dimension

↓

Neural Network

10. Computer-Vision Model

The computer-vision component uses a ResNet-18 convolutional neural network architecture.

The model is created using the timm library.

The application does not use pretrained ImageNet weights during inference. Instead, it creates a ResNet-18 architecture with the number of output classes determined by the classes stored in the trained checkpoint.

The trained model parameters are loaded from:

injury_model.pth

The checkpoint contains:

model_state

classes

The model_state contains the learned neural-network parameters.

The classes value provides the injury categories recognized by the model.

11. Model Loading

The trained model is loaded using PyTorch.

The checkpoint is loaded with:

map_location="cpu"

This allows the prototype to run inference on a computer without requiring a dedicated GPU.

After loading the trained parameters, the model is placed into evaluation mode.

This ensures that the model operates as an inference model rather than performing training operations.

12. Prediction Process

When the user selects:

RUN AI ASSESSMENT

the uploaded image is transformed into the required tensor representation.

The tensor is then passed through the injury-classification model.

The neural network produces a vector of output values corresponding to the possible injury classes.

These raw outputs are converted into probabilities using the softmax function.

Conceptually:

Model Output

↓

Softmax

↓

Class Probabilities

↓

Highest Probability

↓

Predicted Injury

The class with the highest probability is selected as the model's predicted injury.

13. Confidence Calculation

The highest probability produced by the softmax function is displayed as the model's confidence.

For example:

Abrasion 0.82

Burn 0.07

Cut 0.08

Laceration 0.03

The system would report:

Predicted Finding: Abrasion

Confidence: 82%

The application also displays the complete probability distribution as a bar chart.

This allows the user to see not only the selected prediction but also the model's relative probabilities for other classes.

14. Risk-Model Development and Integration

A separate machine-learning model is loaded from:

risk_model.pkl

The model is loaded using joblib.

The application constructs a single patient-data row containing:

age

heart_rate

systolic_bp

diastolic_bp

spo2

respiratory_rate

gcs

pain

The data are passed to the model for prediction.

The model generates:

• A predicted risk class.
• Probabilities for the available risk classes.

The application obtains these values through the model's:

predict()

predict_proba()

methods.

15. Risk-Output Display

The prototype displays the predicted risk category using three visual states:

High → Red

Moderate → Yellow

Low → Green

The probabilities for each available risk class are also displayed.

For example:

High: 15%

Moderate: 25%

Low: 60%

These values represent the output of the trained prototype model.

Because the risk model was trained using synthetic data, the output is explicitly described in the application as a prototype risk estimate rather than a validated clinical risk assessment.

16. Rule-Based Screening Layer

In addition to the machine-learning models, the system contains a simple rule-based screening layer.

The prototype checks several patient variables.

The implemented conditions include:

Heart rate > 120

Systolic BP < 90

SpO₂ < 92

Respiratory rate > 28

GCS < 13

Pain ≥ 8

When a threshold is exceeded or falls below the specified value, the system adds a corresponding screening finding.

For example:

Heart rate > 120

↓

Prototype screening flag

This component is intentionally transparent: unlike the neural network, the conditions are explicitly defined in the software.

17. Findings Module

The findings module combines the image-classification result with the rule-based screening information.

The first finding reports the image classifier's prediction and confidence.

Additional findings are generated if the patient's input values meet any of the prototype screening thresholds.

If no additional thresholds are triggered, the application reports that no additional prototype screening flags were triggered.

This creates a simple summary of the information processed by the application.

18. Protocol Module

The application includes a separate protocol-information component:

protocols.py

The predicted injury class is passed to:

get_protocol(predicted_class)

The function returns protocol-related information including:

title

considerations

The application displays the title and individual considerations.

The application also displays:

PROTOCOL_SOURCE

PROTOCOL_NOTE

This design separates protocol information from the machine-learning model itself.

Therefore, protocol content can be updated independently from the neural-network model.

19. Caching and Performance

The application uses Streamlit's resource caching mechanism:

@st.cache_resource

This is applied to the model-loading functions.

The purpose is to prevent the trained models from being repeatedly loaded every time Streamlit reruns the application.

This improves application responsiveness and reduces unnecessary model-loading operations.

20. Safety and Limitations by Design

Several limitations are explicitly incorporated into the software.

The application states that:

• The system is an educational/research prototype.
• The computer-vision model was trained on a limited wound-image dataset.
• Validation accuracy does not constitute clinical validation.
• The risk model was trained using synthetic data.
• Model confidence does not represent clinical certainty.
• Protocol considerations must be checked against current local guidance.

These statements are included to distinguish software/model performance from clinical validation.

21. Development Testing

After development, the system was evaluated using a test set of 100 injury images.

The evaluation compared the AI's predicted injury classification with the reference classification.

The results were:

Measurement	Result
Images tested	100
Correct classifications	94
Incorrect classifications	6
Overall accuracy	94%
Highest-confidence prediction correct	90%

The model therefore achieved an observed classification accuracy of:

94%

on the evaluated 100-image dataset.

22. Interpretation of Development Results

The 94% result demonstrates the performance of the implemented model on the particular test dataset used in the evaluation.

It does not demonstrate that the system will achieve 94% accuracy on all injury photographs.

The difference between the 94% classification accuracy and the 90% highest-confidence correctness rate also indicates that the model's confidence output should not be treated as equivalent to diagnostic certainty.

The six incorrectly classified images should be retained for error analysis rather than removed from the evaluation.

23. Development Workflow

The overall development process can be summarized as follows:

1. Define project objective

↓

2. Obtain/prepare injury image data

↓

3. Develop/train injury classification model

↓

4. Save trained model checkpoint

↓

5. Develop patient-input interface

↓

6. Develop risk model

↓

7. Develop protocol-information module

↓

8. Integrate components using Streamlit

↓

9. Add screening rules and output displays

↓

10. Test complete application

↓

11. Evaluate 100 test images

↓

12. Analyze accuracy and errors

24. Reproducibility

For the experiment to be reproducible, the following should be preserved:

• Model checkpoint
• Model class labels
• Test image dataset
• Reference classifications
• Image preprocessing procedure
• Software version information
• Source code
• Risk model
• Protocol file
• Evaluation results

Any changes to the model, test dataset, preprocessing method, or evaluation procedure should be documented.

25. Scientific Development Summary

The AI EMS Assessment prototype was developed as a modular Python-based system combining computer vision, machine learning, rule-based screening, and protocol information.

A ResNet-18 convolutional neural network was used for injury-image classification. Uploaded images were converted to RGB format, resized to 224 × 224 pixels, converted to tensors, normalized, and processed by the trained network. Softmax probabilities were used to identify the highest-probability injury classification.

A separate machine-learning model processed patient characteristics and vital signs to generate a prototype risk estimate. A rule-based layer independently identified predefined screening thresholds. Finally, a protocol module associated the predicted injury category with informational protocol considerations.

The integrated system was implemented using Streamlit to provide an interactive interface for entering patient information, uploading images, and viewing model outputs.

In an initial evaluation using 100 injury images, the injury classifier correctly identified 94 images (94%). The highest-confidence prediction corresponded to the reference classification in 90 cases (90%).

These results provide an initial technical evaluation of the prototype. They do not establish clinical validity, clinical safety, or effectiveness. Further testing with larger, independent, and appropriately representative datasets would be required before drawing conclusions about real-world clinical performance.

1. Study Title

Scientific Evaluation of an Artificial Intelligence Computer-Vision Prototype for Injury Classification

2. Purpose of the Study

The purpose of this study is to evaluate the performance of an artificial intelligence (AI) computer-vision prototype designed to classify injuries from photographs.

The evaluation will measure:

• How accurately the AI identifies injuries.
• How frequently the model's highest-confidence prediction is correct.
• What types of errors the model makes.
• Whether model performance varies between injury categories.
• Whether the model's confidence values appear consistent with prediction correctness.

The system is intended as an educational and research prototype. The study does not establish clinical validity, clinical safety, or suitability for diagnosis or treatment.

3. Research Question

Primary Research Question

How accurately can the AI EMS prototype classify injuries from standardized injury photographs?

Secondary Research Question

How reliably does the model's highest-confidence prediction correspond to the correct injury classification?

4. Hypotheses

Hypothesis 1

The AI injury-classification model will correctly classify a high proportion of images in the independent test dataset.

Hypothesis 2

The prediction assigned the highest confidence probability will frequently correspond to the correct injury classification.

Hypothesis 3

The model's performance will vary between injury categories because some injuries may have visually similar characteristics.

5. Experimental Design

The study uses a retrospective image-classification evaluation.

A total of 100 injury images will be evaluated by the AI model.

Each image will have an established reference classification. The reference classification will be treated as the comparison standard for evaluating the model's prediction.

The AI will analyze each image independently and generate:

• Predicted injury category
• Probability for each available injury category
• Highest-confidence prediction
• Confidence percentage

The AI prediction will then be compared with the reference classification.

6. Test Dataset

The evaluation dataset should contain 100 images.

Whenever possible, the test images should be independent from the images used to train the model.

For each image, create a record containing:

Variable	Description
Image ID	Unique identifier
Reference diagnosis	Expected injury classification
AI prediction	Injury classification generated by AI
AI confidence	Probability associated with selected prediction
Correct/incorrect	Whether AI prediction matched reference
Highest-confidence correct	Whether highest-probability prediction matched reference

Example:

Image	Reference	AI Prediction	Confidence	Correct
001	Abrasion	Abrasion	91%	Yes
002	Burn	Burn	84%	Yes
003	Laceration	Cut	73%	No

7. Equipment and Software

The experiment requires:

• Computer capable of running Python
• Python environment
• Streamlit
• PyTorch
• torchvision
• timm
• pandas
• Pillow
• joblib
• Trained injury-classification model
• Test image dataset
• Data-recording spreadsheet or CSV file

The prototype consists of a Streamlit interface containing the computer-vision model, prototype risk model, and protocol-information component.

8. Experimental Procedure

Step 1 — Prepare the Dataset

Place the 100 test images in a dedicated test-data folder.

The images should not be modified during the evaluation unless the same preprocessing procedure is applied consistently to every image.

Assign every image a unique identifier.

For example:

IMG001

IMG002

IMG003

...

IMG100

Record the reference injury classification for every image.

Step 2 — Start the AI System

Launch the Streamlit application.

The application should load:

injury_model.pth

risk_model.pkl

protocols.py

Verify that the application loads without errors.

Step 3 — Establish Standardized Settings

For the image-classification experiment, keep the image-analysis procedure consistent.

The application uses the following preprocessing:

Resize image → 224 × 224 pixels

Convert to RGB

Convert to tensor

Normalize using ImageNet mean and standard deviation

Do not change model settings during the experiment.

Step 4 — Analyze Each Image

Upload the first test image.

Record:

• Image ID
• Reference injury
• AI predicted injury
• Confidence percentage
• Whether prediction was correct

Repeat the procedure for all 100 images.

The evaluator should avoid changing the image or model settings based on the result of an earlier image.

9. Recording Results

Create a spreadsheet with the following columns:

Image_ID

Reference_Injury

AI_Prediction

Confidence

Correct

For example:

Image ID	Reference Injury	AI Prediction	Confidence	Correct
IMG001	Abrasion	Abrasion	91%	Yes
IMG002	Burn	Burn	82%	Yes
IMG003	Cut	Laceration	68%	No

For a more detailed study, also record the probability assigned to every class.

For example:

Image	Abrasion	Burn	Cut	Laceration
IMG001	91%	2%	5%	2%

This allows later analysis of the model's confidence calibration.

10. Primary Outcome Measure

The primary outcome is classification accuracy.

Calculate:

Accuracy = Number of correct predictions ÷ Total number of images

For the completed experiment:

94 correct ÷ 100 images = 0.94

Therefore:

Classification accuracy = 94%

The corresponding error rate is:

6 incorrect ÷ 100 images = 6%

Therefore:

Classification error rate = 6%

11. Confidence Analysis

For every image, identify the class with the highest probability.

Determine whether that class corresponds to the reference diagnosis.

In the completed evaluation:

Highest-confidence prediction correct:

90 / 100 = 90%

Therefore:

Highest-confidence prediction accuracy = 90%

The remaining:

10 / 100 = 10%

represent cases in which the highest-confidence prediction did not correspond to the reference classification.

12. Error Analysis

The six incorrectly classified images should be examined separately.

Create an error table:

Image	Reference	AI Prediction	Confidence	Possible Factor
IMG___	___	___	___%	Image quality
IMG___	___	___	___%	Similar appearance
IMG___	___	___	___%	Lighting
IMG___	___	___	___%	Other

Possible factors should be recorded as observations rather than assumptions.

Examples include:

• Poor lighting
• Blurry photograph
• Obstructed injury
• Unusual injury appearance
• Similar appearance between two classes
• Image containing multiple findings
• Unusual camera angle
• Low resolution

Do not assign a cause unless there is evidence supporting it.

13. Per-Class Analysis

Overall accuracy can hide differences between injury categories.

For example, suppose the dataset contains:

25 abrasions

25 burns

25 cuts

25 lacerations

Calculate accuracy separately for each category.

Example:

Injury Class	Images	Correct	Accuracy
Abrasion	25	24	96%
Burn	25	22	88%
Cut	25	24	96%
Laceration	25	24	96%

These numbers are only an example; the actual values must come from your experiment.

14. Confusion Matrix

A confusion matrix should be created if the model contains multiple injury classes.

For example:

Actual \ Predicted	Abrasion	Burn	Cut	Laceration
Abrasion	—	—	—	—
Burn	—	—	—	—
Cut	—	—	—	—
Laceration	—	—	—	—

The diagonal represents correct classifications.

The off-diagonal cells represent errors.

This can reveal whether the model frequently confuses particular injury categories.

15. Additional Performance Measures

If sufficient data are available, calculate:

Precision

Precision describes how often predictions of a particular class were correct.

Recall

Recall describes how frequently images belonging to a particular class were successfully identified.

F1 Score

The F1 score combines precision and recall into a single measure.

These measures should preferably be reported for each injury category, rather than only as a single overall value.

16. Statistical Analysis

For the 100-image evaluation, the observed classification accuracy was:

94/100 = 94%

A 95% confidence interval can also be calculated for the observed proportion. Using a binomial proportion interval, the approximate 95% confidence interval is 87.5%–97.5%.

This interval describes statistical uncertainty associated with the 100-image sample. It does not mean that the model is guaranteed to achieve an accuracy within this range on future datasets.

The size and composition of the dataset should therefore be considered when interpreting the result.

17. Results

The completed evaluation produced the following results:

Measurement	Result
Images tested	100
Correct classifications	94
Incorrect classifications	6
Classification accuracy	94%
Highest-confidence predictions correct	90
Highest-confidence prediction accuracy	90%

The model correctly classified 94 of the 100 evaluated injury images.

The model's highest-confidence prediction corresponded to the reference classification in 90 of the 100 cases.

18. Discussion

The evaluation demonstrated that the AI prototype correctly classified 94% of the images in the tested dataset.

This result indicates that the model was able to identify the reference injury classification for most images in this particular test set.

However, the six incorrect classifications demonstrate that the model does not classify every image correctly. Examination of these errors is therefore important for understanding the limitations of the system.

The highest-confidence prediction was correct in 90% of cases. Because this percentage was lower than the overall classification accuracy, the confidence values did not perfectly correspond with prediction correctness.

This finding is important because a confidence value generated by an AI model should not automatically be interpreted as clinical certainty.

The results also cannot establish that the model will achieve the same performance on other datasets. Model performance may change depending on image source, image quality, injury category, population, camera, lighting, and other factors.

19. Limitations

Several limitations should be considered.

Sample Size

Only 100 images were evaluated. A larger test dataset would provide a more precise estimate of model performance.

Dataset Representation

The performance of the model depends on whether the test images adequately represent the types of injuries and image conditions encountered in the intended application.

Ground-Truth Classification

The validity of the evaluation depends on the accuracy of the reference classifications.

Model Generalization

Performance on the test dataset does not necessarily predict performance on images from different sources.

Confidence

The model's confidence percentage does not necessarily represent the actual probability that the prediction is clinically correct.

Clinical Validation

This experiment evaluates an AI prototype. It does not constitute clinical validation, and the results should not be interpreted as demonstrating that the system is safe or appropriate for clinical diagnosis or treatment.

20. Future Improvements

Future studies could improve the evaluation by:

• Increasing the number of test images.
• Using an independently collected test dataset.
• Ensuring balanced representation of injury categories.
• Reporting precision, recall, and F1 score for each class.
• Creating a confusion matrix.
• Evaluating model confidence calibration.
• Testing images under different lighting and image-quality conditions.
• Testing images from different cameras and sources.
• Repeating the experiment on a separate dataset.
• Conducting appropriate human-participant research if evaluating usability with EMS students or professionals.

21. Scientific Conclusion

The evaluation of 100 injury photographs showed that the AI EMS prototype correctly classified 94% of the tested images. The model's highest-confidence prediction corresponded to the reference classification in 90% of cases.

These findings provide preliminary evidence of the prototype's performance on the evaluated dataset. However, the results should be interpreted as a technical evaluation of an experimental AI system, rather than evidence of clinical diagnostic effectiveness.

Further evaluation using larger, independent, and more representative datasets would be necessary to determine how consistently the system performs under different conditions.

22. Data Integrity Statement

All results should be recorded before conclusions are drawn. Test images should not be removed because the model classified them incorrectly. Incorrect classifications are part of the experimental result and should be retained in the final analysis.

The final reported accuracy should be calculated from the complete predefined test dataset.

Any changes to the dataset, model, preprocessing procedure, or evaluation criteria should be documented.

23. Recommended Final Figure

Create a bar graph titled:

Figure 1. AI Injury Classification Performance

Use two bars:

Classification Accuracy 94%

Highest-Confidence Accuracy 90%

Label the y-axis:

Percentage Correct (%)

and the x-axis:

Evaluation Measure

A second graph can show:

Correct classifications 94

Incorrect classifications 6

This provides a straightforward visual representation of the experimental results.

24. Final Study Summary

Test dataset: 100 injury images

Correct injury classifications: 94

Incorrect injury classifications: 6

Overall classification accuracy: 94%

Highest-confidence prediction correct: 90%

Highest-confidence prediction incorrect: 10%

Primary finding: The prototype correctly classified 94% of the images in the evaluated dataset.

Interpretation: The result demonstrates performance within this specific experimental dataset but does not establish clinical validity or clinical effectiveness.

Abstract

Emergency medical services (EMS) personnel must rapidly assess patients, identify potentially significant injuries, and determine appropriate next steps in environments where time and information may be limited. Artificial intelligence (AI) has the potential to provide additional informational support during this process through computer vision and machine-learning technologies.

This project developed an AI EMS Assessment prototype combining computer vision, a machine-learning risk model, rule-based screening criteria, and protocol-based informational support. The computer-vision component analyzes uploaded injury photographs and generates an injury classification and probability distribution. The risk component processes patient characteristics and vital signs to generate a prototype risk estimate. The system then displays relevant screening findings and protocol-based considerations through an interactive interface.

In an evaluation of 100 injury images, the injury-classification model correctly classified 94 images, corresponding to an observed accuracy of 94%. The highest-confidence prediction corresponded to the reference classification in 90% of cases. These results provide preliminary evidence of technical performance within the evaluated dataset.

Clinically, a system of this type could potentially be used as an adjunctive decision-support and documentation tool, rather than as a replacement for EMS clinical assessment, medical direction, or established protocols. Potential applications include assisting with structured assessment, highlighting predefined abnormalities, organizing information, and providing informational support. Before clinical deployment, substantial additional validation would be required, including prospective clinical testing, external validation, safety evaluation, bias assessment, human-factors testing, cybersecurity and privacy assessment, regulatory review, and appropriate clinical oversight.

1. Background

Emergency medical services operate in environments that can be unpredictable, time-sensitive, and information-limited. EMS personnel may need to evaluate injuries while simultaneously collecting vital signs, obtaining a patient history, documenting findings, and determining appropriate care according to applicable protocols.

Computer vision is a branch of artificial intelligence that allows computers to process and classify visual information. In healthcare research, computer-vision systems have been investigated for applications involving medical images and other forms of visual information.

Machine-learning systems can also process structured variables such as age and vital signs. When combined with an appropriate user interface, these technologies could potentially provide EMS personnel with an additional source of structured information during an assessment.

The AI EMS Assessment prototype was developed to explore this concept. The system combines injury-image classification with patient information and predefined screening criteria in a single interface.

Importantly, the prototype is not intended to independently diagnose patients or determine treatment. Its proposed clinical role would be limited to providing additional information that a qualified healthcare professional could consider alongside direct patient assessment, established EMS protocols, and medical direction.

2. Problem Statement

A major challenge in emergency assessment is the rapid organization and interpretation of multiple sources of patient information.

For an injury patient, relevant information may include:

• Visual characteristics of the injury
• Vital signs
• Level of consciousness
• Pain severity
• Patient age
• Mechanism and history of injury
• Changes in patient condition

A software system capable of organizing these inputs could potentially reduce the amount of information that must be manually consolidated during an assessment.

However, incorrect AI outputs could also introduce risk. Therefore, any potential clinical application must ensure that AI outputs remain subordinate to professional assessment and established clinical procedures.

3. Proposed Clinical Application

The proposed clinical use of the AI EMS Assessment system would be as an adjunctive informational tool.

A potential workflow could be:

Patient Encounter

↓

EMS Clinical Assessment

↓

Patient Information + Vital Signs

↓

Optional Injury Photograph

↓

AI Processing

↓

Injury Classification

+

Prototype Risk Estimate

+

Screening Flags

+

Protocol Information

↓

EMS Professional Reviews Output

↓

Clinical Assessment + Applicable Protocols

↓

Clinical Decision

The AI would therefore function as an additional information layer rather than the final decision-maker.

4. Potential EMS Use

4.1 Injury Documentation

An injury photograph could potentially be analyzed and associated with a classification generated by the computer-vision model.

The resulting information could potentially assist with organizing documentation.

For example, the system could display:

Predicted finding: Laceration

Model confidence: 82%

The classification would remain an AI-generated observation and would require professional interpretation.

4.2 Structured Patient Assessment

The system could provide a centralized location for entering:

• Heart rate
• Blood pressure
• SpO₂
• Respiratory rate
• GCS
• Pain
• Age

Displaying these variables together could help organize information during a simulated or potentially future clinical assessment.

4.3 Screening Flags

The prototype contains predefined screening thresholds.

For example, the system currently identifies:

Heart rate > 120

Systolic BP < 90

SpO₂ < 92%

Respiratory rate > 28

GCS < 13

Pain ≥ 8

When a threshold is reached, the system displays a screening flag.

In a potential clinical implementation, such flags could be used to draw attention to recorded values that warrant further assessment.

They should not independently determine patient disposition or treatment.

5. Potential Use as Decision Support

A future validated version could potentially provide a structured summary such as:

Injury classification:

[AI-generated finding]

Vital signs:

HR: ___

BP: ___

SpO₂: ___

RR: ___

GCS: ___

Screening findings:

[Detected findings]

Protocol information:

[Applicable information]

This could potentially reduce information fragmentation by presenting multiple assessment components in one interface.

The final interpretation would remain the responsibility of the qualified healthcare professional.

6. Protocol Integration

The prototype includes a protocol-information component that associates injury classifications with predefined considerations.

A future clinical implementation could potentially connect this component to an institution's current, approved protocols.

However, protocol content can vary between jurisdictions and organizations. Therefore, a clinical system should not assume that a single set of recommendations applies universally.

A production system would require:

• Clearly identified protocol source
• Version information
• Effective dates
• Controlled updates
• Local medical-director approval where applicable
• Mechanisms to prevent outdated protocol information from being presented as current

The AI system should not create or modify medical protocols autonomously.

7. Proposed Role of Healthcare Professionals

The proposed system would operate under a human-in-the-loop model.

Under this approach:

• The EMS professional performs the initial patient assessment.
• Relevant information is entered into the system.
• The AI generates informational outputs.
• The EMS professional reviews the outputs.
• The professional compares the outputs with the patient's actual condition and applicable protocols.
• Clinical decisions remain under appropriate professional authority.

This approach is important because an AI model can produce an incorrect classification even when its confidence is high.

8. Potential Clinical Benefits

If appropriately validated, an AI system of this type could potentially provide several benefits.

Information organization

Patient information, image findings, and screening results could be presented in a standardized format.

Additional informational support

The system could provide another computational assessment of an injury photograph.

Consistency

A standardized interface could encourage consistent recording of selected patient variables.

Documentation support

Image classifications and structured patient information could potentially contribute to organized documentation.

Educational applications

The system could also be used in simulation and EMS education to demonstrate how AI-generated information could interact with patient assessment.

These potential benefits remain hypothetical until demonstrated through appropriate clinical research.

9. Risks and Clinical Limitations

Clinical implementation would introduce important risks.

Incorrect classification

The computer-vision model may incorrectly classify an injury.

In the current 100-image evaluation, six images were incorrectly classified.

False reassurance

A high model confidence could potentially cause a user to place excessive trust in an incorrect prediction.

Missing clinical context

A photograph cannot capture every clinically relevant characteristic of an injury or patient.

Dataset bias

If training and testing data do not adequately represent the intended population and injury types, performance may differ in real-world settings.

Model drift

Performance could change over time as injury patterns, imaging devices, patient populations, or operational environments change.

Protocol changes

Clinical guidance can change. A system containing outdated protocol information could potentially provide inappropriate information.

10. Privacy and Security

A clinical version of the system would need strong privacy and security protections.

Injury photographs can contain identifiable patient information. Therefore, a clinical implementation would need appropriate safeguards for:

• Image storage
• Patient identifiers
• Data transmission
• Access control
• Authentication
• Audit logs
• Data retention
• Data deletion
• Encryption

The current prototype should not be assumed to meet clinical privacy or cybersecurity requirements.

11. Clinical Validation Requirements

The current 100-image evaluation is a technical evaluation and should not be considered clinical validation.

Before clinical deployment, additional studies would be needed.

These could include:

Retrospective validation

Evaluate the system using a large, independently collected dataset.

External validation

Evaluate the model using data from locations and sources not represented in model development.

Prospective evaluation

Evaluate system performance in a real-world or appropriately simulated clinical environment.

Human-factors evaluation

Determine whether users correctly understand and appropriately respond to the system's outputs.

Safety evaluation

Identify circumstances in which incorrect AI outputs could result in harmful decisions.

Bias and subgroup evaluation

Evaluate performance across relevant patient and injury subgroups.

Reliability testing

Determine whether system performance remains stable under different image qualities, devices, lighting conditions, and operational environments.

12. Current Experimental Results

The prototype was evaluated using 100 injury images.

The results were:

Measure	Result
Images evaluated	100
Correct classifications	94
Incorrect classifications	6
Classification accuracy	94%
Highest-confidence prediction correct	90%

The observed accuracy of 94% indicates that the model correctly classified most images in the evaluated dataset.

However, the six incorrect classifications demonstrate that the model can produce errors.

The 90% highest-confidence correctness rate further demonstrates that model confidence does not guarantee correctness.

13. Discussion

The prototype demonstrates the technical feasibility of combining computer vision, structured patient data, machine learning, and protocol information within a single EMS-oriented application.

The 94% observed accuracy in the 100-image evaluation provides an initial measurement of computer-vision performance. However, the result should be interpreted within the specific conditions of the experiment.

The proposed clinical role of the system would therefore be fundamentally different from autonomous diagnosis. Rather than replacing professional assessment, a future validated version could potentially provide an additional layer of structured information.

The distinction between prediction and clinical diagnosis is critical. An AI model analyzes the information provided to it according to patterns learned during development. A healthcare professional evaluates the patient using information that may include physical examination, history, mechanism of injury, vital signs, clinical context, and other factors that may not be available to the AI system.

Consequently, an AI-generated classification should be treated as one source of information rather than as a definitive diagnosis.

14. Future Development

Future development could focus on:

• Increasing the size and diversity of the image dataset.
• Collecting an independent external test dataset.
• Measuring per-class precision, recall, and F1 scores.
• Developing a confusion matrix.
• Evaluating confidence calibration.
• Improving image-quality robustness.
• Testing different cameras and lighting environments.
• Improving protocol-version management.
• Adding comprehensive audit logging.
• Conducting human-factors research.
• Conducting appropriate clinical validation studies.
• Evaluating regulatory requirements before any clinical deployment.

15. Proposed Clinical Implementation Model

If future validation demonstrates acceptable performance and the necessary clinical, regulatory, privacy, and safety requirements are satisfied, the system could potentially be implemented as follows:

PATIENT

↓

EMS Initial Assessment

↓

┌─────────┴─────────┐

↓ ↓

Patient Data Injury Image

↓ ↓

└─────────┬─────────┘

↓

AI SYSTEM

↓

┌────────────┼────────────┐

↓ ↓ ↓

Injury Finding Risk Screening

Estimate Flags

│ │ │

└────────────┼────────────┘

↓

Human Review

↓

Applicable EMS Protocols

↓

Clinical Decision

The key principle is that the AI system remains an assistive layer within the broader clinical workflow.

16. Conclusion

The AI EMS Assessment prototype demonstrates a possible approach for integrating computer vision, machine learning, structured vital-sign information, and protocol-based information into an EMS-oriented software platform.

In the initial evaluation of 100 injury images, the injury classifier achieved an observed accuracy of 94%, while its highest-confidence prediction was correct in 90% of cases.

These results demonstrate preliminary technical performance but do not establish clinical effectiveness.

If further research demonstrates sufficient reliability, safety, generalizability, and usability, a system based on this concept could potentially be developed as an adjunctive tool for EMS assessment, documentation, education, and decision support.

Any future clinical implementation would require appropriate validation, clinical oversight, privacy protections, human-factors evaluation, regulatory consideration, and integration with current local EMS protocols.

The fundamental intended role of the technology should remain supportive rather than autonomous: AI may help organize and interpret information, while qualified healthcare professionals remain responsible for evaluating the patient and making clinical decisions.