Effective learning for children with DS relies heavily on visual and interactive elements, as research has consistently demonstrated that illustrated content enhances comprehension and retention [
Recognizing these cognitive needs, Risk Resist was specifically designed to ensure that children with DS could absorb and apply emergency response strategies effectively. Studies highlight that many children with DS struggle with basic life skills [
Risk Resist is a GBL tool designed to teach crisis and emergency management in a safe, controlled environment, preventing real-world risks to children with DS. Through engaging gameplay, players can practice methodical crisis response, such as handling indoor fires, in a way that reinforces learning without exposure to danger. The game’s systemic framework allows for broad application across different types of emergencies, spanning both human-caused and ecological crises. The game is structured into two main categories: (1) human-caused crises includes electrical fires, first aid emergencies (such as fainting and bleeding), and road-crossing safety; and (2) natural catastrophes teaches essential survival strategies during disasters like earthquakes.
To ensure accessibility for children with DS, Risk Resist incorporates specialist recommendations and evidence-based learning theories [
Visual learning: Vivid graphics, animations, and interactive storytelling engage players and reinforce emergency concepts.
Repetition and feedback: Scenarios provide repeated exposure and immediate positive reinforcement, enhancing skill retention.
Simplified instructions: Clear guidance, intuitive controls, and step-by-step instructions ensure accessibility and comprehension for children with DS.
The game’s homepage emphasizes peer support, fostering collaboration and engagement among students. The mechanics implement ongoing trials, ensuring players grasp key emergency responses through adaptive learning techniques essential for children with cognitive difficulties [
Constructivist learning theory: Active learning where individuals acquire knowledge through experiential engagement rather than passive instruction. This aligns with Risk Resist’s interactive gameplay, which enables children to practice crisis management in a simulated environment, reinforcing problem-solving and decision-making skills [
Skill transfer model: The concept of learning transfer refers to the ability to apply knowledge gained in one environment to real-world situations. Research indicates that learning transfer improves when activities mimic real-world challenges [
Reinforcement theory: The game incorporates positive reinforcement by providing instant feedback, reward mechanisms, and gradual difficulty progression. According to Skinner’s reinforcement theory [
Prototype creation and iterative design are described in detail in
The Galaxy Shop.
Risk Resist is built using the Unity 3D engine, ensuring smooth performance and an optimized user experience. The user interface is available with instructions provided in Arabic, catering to Arabic-speaking children with DS and enhancing inclusivity. By combining educational principles, cognitive accessibility, and dynamic engagement strategies, Risk Resist stands out as an effective and innovative tool for enhancing emergency preparedness in children with DS, supporting both their learning and personal development.
One significant issue in many traditional games is the static difficulty of each level [
DDA automatically adjusts scenarios, parameters, and game behaviors in real time based on the player’s skill level. This ensures that the game maintains an optimal balance between challenge and engagement, avoiding boredom (when tasks are too easy) or frustration (when tasks are too difficult). By comparing the player’s current performance to an ideal or reference performance (such as the average first-time performance on a level), the game dynamically modifies difficulty settings.
For children with DS, the DDA algorithm incorporates additional behavioral data points, such as response times, success rates, and specific patterns of behavior. These factors help tailor the game to the child’s unique skill level, ensuring an experience that is neither overwhelming nor trivial.
To simplify understanding, the implementation steps are outlined as follows:
The process of initialization is as follows:
Reference baseline: The algorithm starts with a predefined baseline performance measure (eg, the average number of attempts taken by a typical player to complete a level).
Adjustment rate: An adjustment rate is defined to control how quickly the difficulty level changes based on the player’s performance.
Initial difficulty: The initial difficulty level is set at 0.5, representing a medium difficulty on a scale from 0.0=easiest to 1.0=hardest.
The player’s performance is calculated using factors such as:
Number of attempts to complete the task: how many tries the player needs to succeed.
Time taken to complete the task: longer times suggest increased difficulty.
Success rate (percentage of successful attempts): tracks completion accuracy and failure patterns.
Average response time: indicates reaction speed and cognitive engagement.
Difficulty level adjustment:
The difficulty level is updated by comparing the player’s performance to the reference baseline.
It is normalized based on the time taken, with lower difficulty levels indicating easier tasks.
Ease level is defined as the complement of the difficulty level and reflects how easy the game feels for the player.
Since the machine learning model (Random Forest Regressor) requires data accumulation before making accurate predictions, the system uses simplified difficulty adjustment logic in the early stages: if the player struggles (low success rate and long response times), the difficulty decreases to enhance accessibility; if the player performs well (high success rate and fast response times), the difficulty increases to provide a greater challenge; and difficulty is constrained within 0.0 to 1.0, ensuring balanced play.
This fallback mechanism prevents extreme difficulty spikes or drops before enough data have been collected for machine learning–based adjustments.
Once the player completes at least 5 gameplay interactions, the collected performance data are stored to train the Random Forest Regressor, which identifies patterns in player performance, predicts an optimal difficulty level for upcoming gameplay, and gradually replaces the initial simplified logic with machine learning–driven difficulty optimization.
Once the model is trained, difficulty modifications become more precise, ensuring real-time adaptation based on deeper behavioral analysis. The algorithm predicts the ideal difficulty based on player history, refines the adjustment process using progressively more performance data, and maintains engagement levels while preventing frustration and boredom.
Justification for premodel adjustment logic is present in
Before the Random Forest Regressor gathers enough data for predictions, the system applies a simplified manual difficulty adjustment mechanism to ensure that players experience appropriate challenge levels from the start.
This logic determines whether difficulty should increase or decrease based on immediate performance metrics such as number of attempts, success rate, and response time.
If fewer than 5 gameplay interactions are available, difficulty changes based on predefined rules, preventing extreme difficulty spikes before the machine learning model starts making predictions.
Once the player completes at least 5 gameplay interactions, sufficient performance data have been collected, allowing the Random Forest Regressor to train.
The machine learning model replaces manual logic, making predictions based on the child’s past performance trends.
Unlike early-stage adjustments, these predictions continuously refine difficulty adaptation, ensuring optimal challenge levels by detecting deeper behavioral patterns.
To personalize and optimize the gaming experience for children with DS, we incorporate additional behavioral data points such as response times, success rates, and specific patterns of behavior. These data points are critical for several reasons (
Significance: Response time is a valuable indicator of a player’s processing speed and reaction ability. For children with Down syndrome, who may have varying cognitive and motor skills, measuring response times helps in understanding how quickly they can process and act on game stimuli [
Use: By analyzing response times, the dynamic difficulty adjustment algorithm can dynamically adjust the game pace to match the player’s capabilities, ensuring that tasks are neither too fast nor too slow, thus maintaining an optimal challenge.
Significance: The success rate reflects the player’s ability to complete tasks accurately. It provides insight into how well the player understands and executes game objectives [
Use: Monitoring success rates allows the algorithm to identify when a player is consistently succeeding or struggling. This information helps in adjusting the difficulty level to provide a balanced experience that fosters learning and skill development without causing frustration or disengagement.
Significance: Specific patterns of behavior, such as repeated mistakes or preferred strategies, offer a deeper understanding of the player’s learning style and problem-solving approach [
Use: By identifying these patterns, the dynamic difficulty adjustment algorithm can tailor game scenarios to align with the player’s learning preferences, providing personalized support and encouragement. This ensures that the game remains engaging and educationally effective.
These behavioral data points are used to dynamically adjust the game’s difficulty in real time. The use of a Random Forest Regressor, a sophisticated machine learning model, allows the algorithm to predict the optimal difficulty level based on the collected performance data. By continuously adapting the challenge to the player’s skill level, the DDA algorithm ensures that the game remains appropriately challenging and engaging. This approach prevents boredom from tasks that are too easy and frustration from tasks that are too difficult, ultimately enhancing the learning and development of children with DS.
The Random Forest Regressor was trained using performance data from gameplay sessions. The input features included average response time, success rate, number of attempts, and error frequency. The target variable was the difficulty score, a continuous value between 0 (easy) and 1 (hard), which represented the ideal challenge level for the next session. A total of 162 valid session logs were collected from the 9 children in the experimental group. To evaluate model performance, we applied 5-fold cross-validation. The following regression metrics were used: mean absolute error (MAE), root-mean-square error (RMSE), and
The pseudocode snippet in
BEGIN DDA Algorithm
//Step I: Initialization
// Set the baseline performance reference
baseline_performance ← average_attempts_to_complete_level
adjustment_rate ← predefined_adjustment_rate
current_difficulty ← 0.5 // Medium difficulty (range: 0.0 to 1.0)
//Step II: Performance Calculation
WHILE game_running DO
// Calculate player performance based on key metrics
player_attempts ← get_number_of_attempts()
time_taken ← get_task_completion_time()
success_rate ← calculate_success_rate(player_attempts)
response_time ← calculate_average_response_time()
// Update difficulty level based on player performance
performance_measure ← player_attempts - baseline_performance
normalized_difficulty ← normalize(performance_measure, time_taken)
current_difficulty ← normalized_difficulty
// Compute ease level (complement of difficulty)
ease_level ← 1.0 - current_difficulty
END WHILE
//Step III: Data Collection
WHILE game_running DO
// Collect and store player performance data
data_collected ← {player_attempts, time_taken, success_rate, response_time, current_difficulty}
store(data_collected)
END WHILE
//Step IV: Model Training
IF collected_data_count > 5 THEN
// Train Random Forest Regressor using collected data
model ← RandomForestRegressor()
model.train(collected_data, difficulty_levels)
END IF
//Step V: Difficulty Adjustment
IF model.is_trained THEN
// Predict difficulty using trained model
predicted_difficulty ← model.predict(current_player_data)
current_difficulty ← clamp(predicted_difficulty, 0.0, 1.0)
ELSE
// Use fallback logic for adjusting difficulty
IF current_difficulty > 0.5 THEN
current_difficulty ← current_difficulty - adjustment_rate // Make easier
ELSE
current_difficulty ← current_difficulty + adjustment_rate // Make harder
END IF
current_difficulty ← clamp(current_difficulty, 0.0, 1.0)
END IF
//Step VI: Implementation
// Use scikit-learn's RandomForestRegressor for model implementation
model.initialize()
model.calculate_performance_measures()
model.train(collected_data)
model.adjust_difficulty(current_difficulty)
END DDA Algorithm
# Initialize difficulty level at a moderate starting point (0.5)
difficulty_level = 0.5
min_difficulty = 0.0
max_difficulty = 1.0
# Function to manually adjust difficulty before ML predictions start
def pre_model_difficulty_adjustment(success_rate, avg_response_time):
global difficulty_level
# Check if the player is struggling
if success_rate < 0.4 or avg_response_time > threshold_time:
difficulty_level -= 0.1 # Decrease difficulty to make tasks easier
elif success_rate > 0.8 and avg_response_time < threshold_time:
difficulty_level += 0.1 # Increase difficulty to make tasks more challenging
# Ensure difficulty remains within valid bounds
difficulty_level = max(min_difficulty, min(difficulty_level, max_difficulty))
return difficulty_level
# Example usage before machine learning model is trained
new_difficulty = pre_model_difficulty_adjustment(player_success_rate, player_avg_response_time)
The DDA algorithm flow diagram. DDA: dynamic difficulty adjustment.
To assess the effectiveness of Risk Resist, a controlled experimental design was used, comparing GBL with traditional human-based training. The study aimed to evaluate learning gains and engagement outcomes, structured around the following hypotheses:
Null hypothesis (hypothesis 0): No significant difference in learning outcomes or engagement between Risk Resist and traditional human-based training.
Hypothesis 1 (learning gain hypothesis): Children who play Risk Resist will demonstrate significantly higher learning gains compared to those receiving traditional human-based training.
Hypothesis 2 (engagement level hypothesis): Children who play Risk Resist will exhibit significantly higher engagement levels compared to those receiving traditional human-based training.
A between-group experimental design was used, with participants divided equally into treatment and control groups (n=9 each). The control group received traditional teacher-led training for the same emergency scenarios presented in the game, while the experimental group engaged in all Risk Resist gameplay sessions. The main group used laptops to engage with Risk Resist, while the control group underwent teacher-led training. The sample consisted of 18 children with DS, aged 8 to 12 years, recruited from special needs facilities and schools. Selection criteria emphasized functional living skills, such as the ability to use the toilet and recognize objects, with the severity of DS being a more critical criterion than age.
To ensure the internal validity of the study and account for the relatively small sample size (n=18), a baseline equivalence test was conducted to verify that the experimental (Risk Resist) and control (traditional training) groups did not differ significantly in their initial emergency preparedness knowledge before the intervention. An independent samples 2-tailed
The study used various experimental materials, including pretest and posttest surveys, participant performance assessments, and parent consent forms. Testing was conducted individually in classrooms over a 2-week period. Teachers participated in the evaluation phase by completing observation-based questionnaires designed to measure learning outcomes and engagement levels.
Participants were given identical pretests and posttests comprising 5 questions covering emergency scenarios (
Teachers were available to assist participants in understanding the test questions. Their assistance was limited to rephrasing or explaining the questions in simpler terms without providing any answers. This ensured that participants understood the instructions while maintaining the integrity of the test.
To evaluate engagement levels, a Mann-Whitney
Learning gain=posttest score–pretest score
A 5-point Likert scale questionnaire evaluated engagement factors like attention, motivation, and effort during gameplay or training. Teachers completed the questionnaire based on their observations of participants. The total engagement score for each participant was derived by summing their responses across all items. Scores were normalized to allow consistent comparison between groups.
Participants were asked 5 questions regarding crisis and emergency scenarios during both the pretest and posttest phases. Learning gains were calculated as the difference between correct answers in the pretest and posttest, as shown in equation 2:
Learning gain=correct answers in the posttest–correct answers in the pretest
The questions asked of the children are presented in
What will you do if there is a fire caused by electricity? What is the difference between a regular fire and a fire caused by electricity?
Can you give quick help to anyone who faints in front of you?
What can you do if you or the other one in front of you has bleeding because of an injury?
Can you cross the road? What do you do while crossing the road?
If there is an earthquake, what will you do? (with a quick explanation what is the meaning of earthquake)
Pilot testing with a small sample of children ensured the reliability and suitability of the game for its target demographic. To assess the effectiveness of the game, learning outcomes were compared between the 2 groups by analyzing the average learning gains of participants in each condition. Engagement levels were measured using a 5-point Likert scale questionnaire [
After the session, teachers were provided hard copies of the questionnaire and instructed to complete them based on their observations of children’s engagement and learning behaviors. An engagement questionnaire presented to the teachers upon their observations of participants (
Engagement questionnaire presented to the teachers upon their observations of participants.
|
|
Question | 1=very poor | 2=poor | 3=fair | 4=good | 5=very good |
| 1. | To what extent did the game or training hold the child’s attention? |
|
|
|
|
|
| 2. | How much effort did the child put into playing the game or getting the training? |
|
|
|
|
|
| 3. | Does the game or training motivate the child to learn? |
|
|
|
|
|
| 4. | The child was motivated to get involved in the training or game. |
|
|
|
|
|
| 5. | I believe the game or training has improved the child’s understanding of the covered topics. |
|
|
|
|
|
| 6. | To what extent the child was tired after playing the game? |
|
|
|
|
|
| 7. | How well does the game maintain the child’s interest over time? |
|
|
|
|
|
| 8. | Was the total duration of the game or training satisfactory to understand the required scenarios? |
|
|
|
|
|
| 9. | Would the child like to play the game or get training again? |
|
|
|
|
|
| 10. | How well do you think the child performed in the game or training? |
|
|
|
|
|
To assess the relationship between learning gains and engagement, a Pearson correlation coefficient (
where
This study involving human participants was conducted following the ethical standards of the institutional review board at the Faculty of Computers and Information, Mansoura University, Egypt (protocol 202110016). Parents of all participating children provided written informed consent before enrollment, authorizing both participation in the study and the subsequent use of data for research and publication purposes. Participation was entirely voluntary, and no monetary or material compensation was offered to participants or their families. To protect participant privacy and confidentiality, all collected data were anonymized and deidentified prior to analysis; no personally identifiable information was stored or shared. Additionally, no identifiable images of participants are included in the manuscript or supplementary materials. The study was conducted following ethical standards outlined in the Declaration of Helsinki.