AI-Powered Fall Risk Prediction in Nursing Care

Type of Research:

This research is planned as Design and Development Research as an innovative system will be developed.

Place and Time of the Research:

The research will be conducted at the Physical Medicine and Rehabilitation department of a research hospital located in eastern Turkey between June and August 2025.

Population and Sample of the Study:

The study population consisted of all patients treated at the Physical Medicine and Rehabilitation Service at Turgut Özal Medical Center. The sample size for the study was determined to be 161 using a power analysis, with a confidence interval of 0.95, a bias level of 0.05, an effect size of 0.30, and a representativeness of 0.95.

Individuals who meet the inclusion criteria will be included in the sample of the study until the sample size is reached by non-probability sampling method.

General Framework:

The development of the fall risk assessment application in this study will follow a structured methodology. It begins with the selection of participants who meet the defined inclusion and exclusion criteria. For each participant, a descriptive information form will be created, including basic data such as age, height, weight, and other relevant variables. This information will be evaluated in conjunction with the Morse Fall Scale, which will be used to assess each participant's risk of falling. Patients included in the sample will first be assessed using the Morse Fall Scale, followed by a gait analysis. This gait assessment will be conducted using a computer vision system that employs a digital camera to monitor participants as they walk at a comfortable pace in a clinical corridor. Additionally, participants will carry an accelerometer in their pockets to record acceleration data along the X, Y, and Z axes during walking. The collected data will be processed in real time using machine learning techniques to estimate the biomechanics of the lower and upper limbs. Key gait features-such as stride length, cadence, gait period, and range of motion (ROM)-will be extracted. These features, along with Morse Fall Scale scores, will be used to train and validate an artificial neural network (ANN) designed to identify potential correlations between gait characteristics and fall risk levels.

As a result, the final system will be able to predict the fall risk of new cases in real time through gait analysis.

Computer Vision System:

The computer vision system will consist of one camera connected to a computer and will be configured to record at 20 frames per second (FPS). The camera will be mounted on a tripod to ensure stable imaging of all extremities. The system will be complemented by artificial illumination from infrared reflectors with a wavelength of 850 nm to ensure uniform illumination of the study area without disturbing the participants.

Walk Evaluation Protocol:

A biomechanical gait study will be performed using computer vision to determine kinematic parameters. The methodology used to obtain this information requires participants to move through the clinic corridor at their own comfort pace. Patients will also be able to walk with the help of a companion or nurse. They will be informed about this process in advance to ensure that it does not pose any risk to their health or safety. Each test can be interrupted at any time if the participant feels they need assistance due to fatigue, discomfort or dizziness

Biomechanics Prediction:

For image analysis, we will use You Only Look Once, version 8 (YOLOv8) , a high-speed, high-precision model used in computer vision. YOLOv8 supports tasks such as detection, segmentation, pose estimation, tracking and classification. This model can be trained using specialized databases to recognize specific objects such as humans and track key points. The development will be based on the placement of key points to indicate different segments of the left and right extremities. For the development of the algorithm, we have chosen to use the open-source Keras library in Python, designed for the training and implementation of deep learning models. These models can be used both for the development of neural networks for the analysis of statistical databases and for the training of models for the analysis of images performing tasks such as the detection of objects, people, faces and body orientation. Therefore, using the database obtained from Keras, the training of personalized models will be carried out, where information and images are divided for training and validation. The resulting model will then be used for the analysis of gait test videos. Here, each frame will be taken to perform the identification of each limb segment through key points

Machine Learning Based Classifiers:

In this study, four classifiers will be used to classify gait data, including convolutional neural network (CNN), long short-term memory (LSTM) neural network, support vector machine (SVM) and K-nearest neighbor (KNN) classifier. A CNN with two 2D convolutional layers will be used as a reference. CNN is a type of deep neural network classifier used in image classification and recognition. It can recognize features directly from the data instead of extracting them manually. In the CNN structure in this study, the data will be passed through various layers with different tasks. The input will be transferred to the convolution layer where a spatial filter is applied to the inputs in the form of a weight window. This approach has been used in medical applications where the datasets are small in size and has shown success in medical image recognition.

Data Collection:

The data will be collected in the Physical Medicine and Rehabilitation department of a research hospital located in eastern Turkey between June and August 2025. Data will be gathered by the researcher through questionnaires and measurements using the face-to-face interview method.

Data Collection Tools: Descriptive Information Form, Morse Fall Risk Scale, a camera with video capture feature and an accelerometer will be used to collect the data.

Descriptive Information Form: This form was prepared within the scope of the master's thesis titled "Computer Vision in Fall Risk Prediction with Machine Learning: Development of a Nursing Clinical Decision Support System". The aim of our study is to predict the fall risk of patients hospitalized in the neurology clinic with machine learning and computer vision methods.

Morse Falls Scale: Developed in 1985 by Janice M. Morse in a randomized controlled study with 100 falling patients and 100 non-falling patients, the Morse Falls Scale was adapted to Turkish by Demir and Intepeler in 2012. It is an effective and simple measurement tool that is frequently used in hospitals in Turkey and used to diagnose potential patient fall risks for the nursing profession. The scale consists of six criteria (secondary diagnosis, presence of a history of falls, mobilization support, presence of intravenous access or heparin use, gait/transfer, and mental status) that diagnose fall risk. According to this assessment tool, if the patient has a score below 25 points, he/she is in the low risk group for falls. If the score is between 25 and 50, the patient is in the medium risk group, and if the score is 51 and above, the patient is in the high risk group. A minimum score of 0 and a maximum score of 125 can be obtained from the scale. This scale allows a systematic determination of the fall risk of patients in clinical settings. While 82.9% of the nurses stated that the Morse Falls Scale was easy and quick to use, 54% of the nurses stated that they diagnosed potential fall risks by allocating less than three minutes to the patient while using this scale. Although the Cronbach's Alpha value of the original scale was 0.16, the Cronbach's Alpha value in the Turkish validity and reliability study was calculated as 0.55.

Video Recording: The spontaneous gait of the patients will be recorded for 2 minutes with a video recording device to create a database for computer vision and machine learning methods.

Accelerometer: This simple device is sized enough for patients to carry in their pockets and measures the accelerations of movement in the X, Y and Z axes during spontaneous walking. These measurements will be recorded and collected simultaneously with video recording for 2 minutes and will provide a database for the system to be created with the deep learning method.

Data Evaluation:

In this research, One-way Analysis of Variance (one-way ANOVA) with Post-hoc Least Significant Difference (PostHoc LSD) will be used to determine the statistical difference in the average range of motion of a joint between different gait patterns and the average gait pattern recognition accuracy between machine learning algorithms. Pearson chi-square or Bonferroni correction test will be performed to determine the difference in each gait shape recognition accuracy between multiple machine learning algorithms for different gait shapes. RStudio software will be used for statistical analysis. p value less than 0.05 will be considered statistically significant.