RESUMEN

The Otago Exercise Program (OEP) represents a crucial rehabilitation initiative tailored for older adults, aimed at enhancing balance and strength. Despite previous efforts utilizing wearable sensors for OEP recognition, existing studies have exhibited limitations in terms of accuracy and robustness. This study addresses these limitations by employing a single waist-mounted Inertial Measurement Unit (IMU) to recognize OEP exercises among community-dwelling older adults in their daily lives. A cohort of 36 older adults participated in laboratory settings, supplemented by an additional 7 older adults recruited for at-home assessments. The study proposes a Dual-Scale Multi-Stage Temporal Convolutional Network (DS-MS-TCN) designed for two-level sequence-to-sequence classification, incorporating them in one loss function. In the first stage, the model focuses on recognizing each repetition of the exercises (micro labels). Subsequent stages extend the recognition to encompass the complete range of exercises (macro labels). The DS-MS-TCN model surpasses existing state-of-the-art deep learning models, achieving f1-scores exceeding 80% and Intersection over Union (IoU) f1-scores surpassing 60% for all four exercises evaluated. Notably, the model outperforms the prior study utilizing the sliding window technique, eliminating the need for post-processing stages and window size tuning. To our knowledge, we are the first to present a novel perspective on enhancing Human Activity Recognition (HAR) systems through the recognition of each repetition of activities.

RESUMEN

To gain insights into the impact of upper limb (UL) dysfunctions after breast cancer treatment, this study aimed to develop a temporal convolutional neural network (TCN) to detect functional daily UL use in breast cancer survivors using data from a wrist-worn accelerometer. A pre-existing dataset of 10 breast cancer survivors was used that contained raw 3-axis acceleration data and simultaneously recorded video data, captured during four daily life activities. The input of our TCN consists of a 3-axis acceleration sequence with a receptive field of 243 samples. The 4 ResNet TCN blocks perform dilated temporal convolutions with a kernel of size 3 and a dilation rate that increases by a factor of 3 after each iteration. Outcomes of interest were functional UL use (minutes) and percentage UL use. We found strong agreement between the video and predicted data for functional UL use (ICC = 0.975) and moderately strong agreement for %UL use (ICC = 0.794). The TCN model overestimated the functional UL use by 0.71 min and 3.06%. Model performance showed good accuracy, f1, and AUPRC scores (0.875, 0.909, 0.954, respectively). In conclusion, using wrist-worn accelerometer data, the TCN model effectively identified functional UL use in daily life among breast cancer survivors.

Asunto(s)

RESUMEN

Eating speed is an important indicator that has been widely investigated in nutritional studies. The relationship between eating speed and several intake-related problems such as obesity, diabetes, and oral health has received increased attention from researchers. However, existing studies mainly use self-reported questionnaires to obtain participants' eating speed, where they choose options from slow, medium, and fast. Such a non-quantitative method is highly subjective and coarse at the individual level. This study integrates two classical tasks in automated food intake monitoring domain: bite detection and eating episode detection, to advance eating speed measurement in near-free-living environments automatically and objectively. Specifically, a temporal convolutional network combined with a multi-head attention module (TCN-MHA) is developed to detect bites (including eating and drinking gestures) from IMU data. The predicted bite sequences are then clustered into eating episodes. Eating speed is calculated by using the time taken to finish the eating episode to divide the number of bites. To validate the proposed approach on eating speed measurement, a 7-fold cross validation is applied to the self-collected fine-annotated full-day-I (FD-I) dataset, and a holdout experiment is conducted on the full-day-II (FD-II) dataset. The two datasets are collected from 61 participants with a total duration of 513 h, which are publicly available. Experimental results show that the proposed approach achieves a mean absolute percentage error (MAPE) of 0.110 and 0.146 in the FD-I and FD-II datasets, respectively, showcasing the feasibility of automated eating speed measurement in near-free-living environments.

RESUMEN

Freezing of gait (FOG) is an episodic and highly disabling symptom of Parkinson's disease (PD). Although described as a single phenomenon, FOG is heterogeneous and can express as different manifestations, such as trembling in place or complete akinesia. We aimed to analyze the efficacy of deep learning (DL) trained on inertial measurement unit data to classify FOG into both manifestations. We adapted and compared four state-of-the-art FOG detection algorithms for this task and investigated the advantages of incorporating a refinement model to address oversegmentation errors. We evaluated the model's performance in distinguishing between trembling and akinesia, as well as other forms of movement cessation (e.g., stopping and sitting), against gold-standard video annotations. Experiments were conducted on a dataset of eighteen PD patients completing a FOG-provoking protocol in a gait laboratory. Results showed our model achieved an F1 score of 0.78 and segment F1@50 of 0.75 in detecting FOG manifestations. Assessment of FOG severity was strong for trembling (ICC=0.86, [0.66,0.95]) and moderately strong for akinesia (ICC=0.78, [0.51,0.91]). Importantly, our model successfully differentiated FOG from other forms of movement cessation during 360-degree turning-in-place tasks. In conclusion, our study demonstrates that DL can accurately assess different types of FOG manifestations, warranting further investigation in larger and more diverse verification cohorts.

Asunto(s)

RESUMEN

BACKGROUND: Freezing of gait (FOG) is an episodic and highly disabling symptom of Parkinson's Disease (PD). Traditionally, FOG assessment relies on time-consuming visual inspection of camera footage. Therefore, previous studies have proposed portable and automated solutions to annotate FOG. However, automated FOG assessment is challenging due to gait variability caused by medication effects and varying FOG-provoking tasks. Moreover, whether automated approaches can differentiate FOG from typical everyday movements, such as volitional stops, remains to be determined. To address these questions, we evaluated an automated FOG assessment model with deep learning (DL) based on inertial measurement units (IMUs). We assessed its performance trained on all standardized FOG-provoking tasks and medication states, as well as on specific tasks and medication states. Furthermore, we examined the effect of adding stopping periods on FOG detection performance. METHODS: Twelve PD patients with self-reported FOG (mean age 69.33 ± 6.02 years) completed a FOG-provoking protocol, including timed-up-and-go and 360-degree turning-in-place tasks in On/Off dopaminergic medication states with/without volitional stopping. IMUs were attached to the pelvis and both sides of the tibia and talus. A temporal convolutional network (TCN) was used to detect FOG episodes. FOG severity was quantified by the percentage of time frozen (%TF) and the number of freezing episodes (#FOG). The agreement between the model-generated outcomes and the gold standard experts' video annotation was assessed by the intra-class correlation coefficient (ICC). RESULTS: For FOG assessment in trials without stopping, the agreement of our model was strong (ICC (%TF) = 0.92 [0.68, 0.98]; ICC(#FOG) = 0.95 [0.72, 0.99]). Models trained on a specific FOG-provoking task could not generalize to unseen tasks, while models trained on a specific medication state could generalize to unseen states. For assessment in trials with stopping, the agreement of our model was moderately strong (ICC (%TF) = 0.95 [0.73, 0.99]; ICC (#FOG) = 0.79 [0.46, 0.94]), but only when stopping was included in the training data. CONCLUSION: A TCN trained on IMU signals allows valid FOG assessment in trials with/without stops containing different medication states and FOG-provoking tasks. These results are encouraging and enable future work investigating automated FOG assessment during everyday life.

Asunto(s)

RESUMEN

Otago Exercise Program (OEP) is a rehabilitation program for older adults to improve frailty, sarcopenia, and balance. Accurate monitoring of patient involvement in OEP is challenging, as self-reports (diaries) are often unreliable. The development of wearable sensors and their use in Human Activity Recognition (HAR) systems has lead to a revolution in healthcare. However, the use of such HAR systems for OEP still shows limited performance. The objective of this study is to build an unobtrusive and accurate system to monitor OEP for older adults. Data was collected from 18 older adults wearing a single waist-mounted Inertial Measurement Unit (IMU). Two datasets were recorded, one in a laboratory setting, and one at the homes of the patients. A hierarchical system is proposed with two stages: 1) using a deep learning model to recognize whether the patients are performing OEP or activities of daily life (ADLs) using a 10-minute sliding window; 2) based on stage 1, using a 6-second sliding window to recognize the OEP sub-classes. Results showed that in stage 1, OEP could be recognized with window-wise f1-scores over 0.95 and Intersection-over-Union (IoU) f1-scores over 0.85 for both datasets. In stage 2, for the home scenario, four activities could be recognized with f1-scores over 0.8: ankle plantarflexors, abdominal muscles, knee bends, and sit-to-stand. These results showed the potential of monitoring the compliance of OEP using a single IMU in daily life. Also, some OEP sub-classes are possible to be recognized for further analysis.

Asunto(s)

RESUMEN

Unhealthy dietary habits are considered as the primary cause of various chronic diseases, including obesity and diabetes. The automatic food intake monitoring system has the potential to improve the quality of life (QoL) of people with diet-related diseases through dietary assessment. In this work, we propose a novel contactless radar-based approach for food intake monitoring. Specifically, a Frequency Modulated Continuous Wave (FMCW) radar sensor is employed to recognize fine-grained eating and drinking gestures. The fine-grained eating/drinking gesture contains a series of movements from raising the hand to the mouth until putting away the hand from the mouth. A 3D temporal convolutional network with self-attention (3D-TCN-Att) is developed to detect and segment eating and drinking gestures in meal sessions by processing the Range-Doppler Cube (RD Cube). Unlike previous radar-based research, this work collects data in continuous meal sessions (more realistic scenarios). We create a public dataset comprising 70 meal sessions (4,132 eating gestures and 893 drinking gestures) from 70 participants with a total duration of 1,155 minutes. Four eating styles (fork & knife, chopsticks, spoon, hand) are included in this dataset. To validate the performance of the proposed approach, seven-fold cross-validation method is applied. The 3D-TCN-Att model achieves a segmental F1-score of 0.896 and 0.868 for eating and drinking gestures, respectively. The results of the proposed approach indicate the feasibility of using radar for fine-grained eating and drinking gesture detection and segmentation in meal sessions.

Asunto(s)

RESUMEN

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder hallmarked by challenges in social communication, limited interests, and repetitive, stereotyped movements and behaviors. Numerous research efforts have indicated that individuals with ASD exhibit distinct brain connectivity patterns compared to control groups. However, these investigations, often constrained by small sample sizes, have led to inconsistent results, suggesting both heightened and diminished long-range connectivity within ASD populations. To bolster our analysis and enhance their reliability, we conducted a retrospective study using two different connectivity metrics and employed both traditional statistical methods and machine learning techniques. The concurrent use of statistical analysis and classical machine learning techniques advanced our understanding of model predictions derived from the spectral or connectivity attributes of a subject's EEG signal, while also verifying these predictions. Significantly, the utilization of machine learning methodologies empowered us to identify a unique subgroup of correctly classified children with ASD, defined by the analyzed EEG features. This improved approach is expected to contribute significantly to the existing body of knowledge on ASD and potentially guide personalized treatment strategies.

Asunto(s)

RESUMEN

Objective. Rheumatic Heart Disease (RHD) is one of the highly prevalent heart diseases in developing countries that can affect the pericardium, myocardium, or endocardium. Rheumatic endocarditis is a common RHD variant that gradually deteriorates the normal function of the heart valves. RHD can be diagnosed using standard echocardiography or listened to as a heart murmur using a stethoscope. The electrocardiogram (ECG), on the other hand, is critical in the study and identification of heart rhythms and abnormalities. The effectiveness of ECG to identify distinguishing signs of rheumatic heart problems, however, has not been adequately examined. This study addressed the possible use of ECG recordings for the characterization of problems of the heart in RHD patients.Approach. To this end, an extensive ECG dataset was collected from patients suffering from RHD (PwRHD), and healthy control subjects (HC). Bandpass filtering was used at the preprocessing stage. Each data was then standardized by removing its mean and dividing by its standard deviation. Delineation of the onsets and offsets of waves was performed using KIT-IBT open ECG MATLAB toolbox. PR interval, QRS duration, RR intervals, QT intervals, and QTc intervals were computed for each heartbeat. The median values of the temporal parameters were used to eliminate possible outliers due to missed ECG waves. The data were clustered in different age groups and sex. Another categorization was done based on the time duration since the first RHD diagnosis.Main results. In 47.2% of the cases, a PR elongation was observed, and in 26.4% of the cases, the QRS duration was elongated. QTc was elongated in 44.3% of the cases. It was also observed that 62.2% of the cases had bradycardia.Significance. The end product of this research can lead to new medical devices and services that can screen RHD based on ECG which could somehow assist in the detection and diagnosis of the disease in low-resource settings and alleviate the burden of the disease.

Asunto(s)

RESUMEN

Modelling real-world time series can be challenging in the absence of sufficient data. Limited data in healthcare, can arise for several reasons, namely when the number of subjects is insufficient or the observed time series is irregularly sampled at a very low sampling frequency. This is especially true when attempting to develop personalised models, as there are typically few data points available for training from an individual subject. Furthermore, the need for early prediction (as is often the case in healthcare applications) amplifies the problem of limited availability of data. This article proposes a novel personalised technique that can be learned in the absence of sufficient data for early prediction in time series. Our novelty lies in the development of a subset selection approach to select time series that share temporal similarities with the time series of interest, commonly known as the test time series. Then, a Gaussian processes-based model is learned using the existing test data and the chosen subset to produce personalised predictions for the test subject. We will conduct experiments with univariate and multivariate data from real-world healthcare applications to show that our strategy outperforms the state-of-the-art by around 20%.

Asunto(s)

RESUMEN

Inertial sensors have played a key role in the development of Human Activity Recognition (HAR) systems. Adding gyroscopes in HAR systems leads to increased battery and processing resources. Therefore, it is important to explore their added value compared with using accelerometers only. This study evaluates the added value of gyroscopes in activity recognition. Two public available datasets recorded by accelerometers and gyroscopes were studied. These datasets focus on multiple types of activities: UCI HAR dataset includes walking, walking upstairs, walking downstairs, sitting, standing, laying and WISDM dataset includes 18 hand-oriented and non-hand-oriented activities. Several machine learning models were applied to both datasets for activity recognition. Leave-one-subject-out cross-validation (LOSO) was applied to evaluate the models, where the training set and test set were from different subjects. For UCI HAR dataset, the multilayer perceptron (MLP) model obtained the highest f1-scores. Adding a gyroscope on the waist significantly improved the f1-scores of sitting and laying (both ). For WISDM dataset, the support vector machines (SVM) model obtained the highest f1-scores. The gyroscope on the wrist improved hand-oriented activities while the gyroscope in the pockets improved non-hand-oriented activities (all . The results showed the improvement for recognition performance by adding gyroscopes. However, the improvement was dependent on the type of activity and the mounting place of the gyroscope. Clinical relevance- Gyroscopes are common sensors for activity recognition in wearable healthcare systems. This study proves the added value by adding gyroscopes on different mounting places for recognition performance.

Asunto(s)

RESUMEN

Maintaining adequate hydration is important for health. Inadequate liquid intake can cause dehydration problems. Despite the increasing development of liquid intake monitoring, there are still open challenges in drinking detection under free-living conditions. This paper proposes an automatic liquid intake monitoring system comprised of wrist-worn Inertial Measurement Units (IMU s) to recognize drinking gesture in free-living environments. We build an end-to-end approach for drinking gesture detection by employing a novel multi-stage temporal convolutional network (MS-TCN). Two datasets are collected in this research, one contains 8.9 hours data from 13 participants in semi-controlled environments, the other one contains 45.2 hours data from 7 participants in free-living environments. The Leave-One-Subject-Out (LOSO) evaluation shows that this method achieves a segmental F1-score of 0.943 and 0.900 in the semi-controlled and free-living datasets, respectively. The results also indicate that our approach outperforms the convolutional neural network and long-short-term-memory network combined model (CNN-LSTM) on our datasets. The dataset used in this paper is available at https://github.com/Pituohai/drinking-gesture-dataset/. Clinical Relevance- This automatic liquid intake monitoring system can detect drinking gesture in daily life. It has the potential to be used to record the frequency of drinking water for at-risk elderly or patients in the hospital.

Asunto(s)

RESUMEN

Assessment of physical performance is essential to predict the frailty level of older adults. The modified Physical Performance Test (mPPT) clinically assesses the performance of nine activities: standing balance, chair rising up & down, lifting a book, putting on and taking off a jacket, picking up a coin, turning 360°, walking, going upstairs, and going downstairs. The activity performing duration is the primary evaluation standard. In this study, wearable devices are leveraged to recognize and predict mPPT items' duration automatically. This potentially allows frequent follow up of physical performance, and facilitates more appropriate interventions. Five devices, including accelerometers and gyroscopes, were attached to the waist, wrists and ankles of eight younger adults. The system was experimented within three aspects: machine learning models, sensor placement, and sampling frequencies, to which the non-causal six-stages temporal convolutional network using 6.25 Hz signals from the left wrist and right ankle obtained the optimal performance. The duration prediction error ranged from 0.63±0.29 s (turning 360°) to 8.21±16.41 s (walking). The results suggest the potential for the proposed system in the automatic recognition and segmentation of mPPT items. Future work includes improving the recognition performance of lifting a book and implementing the frailty score prediction.

Asunto(s)

RESUMEN

BACKGROUND: Freezing of gait (FOG) is a common and debilitating gait impairment in Parkinson's disease. Further insight into this phenomenon is hampered by the difficulty to objectively assess FOG. To meet this clinical need, this paper proposes an automated motion-capture-based FOG assessment method driven by a novel deep neural network. METHODS: Automated FOG assessment can be formulated as an action segmentation problem, where temporal models are tasked to recognize and temporally localize the FOG segments in untrimmed motion capture trials. This paper takes a closer look at the performance of state-of-the-art action segmentation models when tasked to automatically assess FOG. Furthermore, a novel deep neural network architecture is proposed that aims to better capture the spatial and temporal dependencies than the state-of-the-art baselines. The proposed network, termed multi-stage spatial-temporal graph convolutional network (MS-GCN), combines the spatial-temporal graph convolutional network (ST-GCN) and the multi-stage temporal convolutional network (MS-TCN). The ST-GCN captures the hierarchical spatial-temporal motion among the joints inherent to motion capture, while the multi-stage component reduces over-segmentation errors by refining the predictions over multiple stages. The proposed model was validated on a dataset of fourteen freezers, fourteen non-freezers, and fourteen healthy control subjects. RESULTS: The experiments indicate that the proposed model outperforms four state-of-the-art baselines. Moreover, FOG outcomes derived from MS-GCN predictions had an excellent (r = 0.93 [0.87, 0.97]) and moderately strong (r = 0.75 [0.55, 0.87]) linear relationship with FOG outcomes derived from manual annotations. CONCLUSIONS: The proposed MS-GCN may provide an automated and objective alternative to labor-intensive clinician-based FOG assessment. Future work is now possible that aims to assess the generalization of MS-GCN to a larger and more varied verification cohort.

Asunto(s)

RESUMEN

During COVID-19, awareness of proper hand washing has increased significantly. It is critical that people learn the correct hand washing techniques and adopt good hand washing habits. Hence, this study proposes using wearable devices to detect hand washing activity among other daily living activities (ADLs) and classify steps proposed by the World Health Organization (WHO). Two experiments were conducted with 16 participants, aged from 20 to 31. The first experiment was hand washing following WHO regulation (ten participants), and the second experiment was performing eight ADLs (eight participants). All participants wore two wearable devices equipped with accelerometers and gyroscopes; one on each wrist. Four machine learning classifiers were compared in classifying hand washing steps in the leave-one-subject-out (LOSO) mode. The SVM model with Gaussian kernel achieved the best performance in classifying 11 washing hands steps, with an average F1-score of 0.8501. When detected among the other ADLs, hand washing following WHO regulation obtained the F1-score of 0.9871. The study demonstrates that wearable devices are feasible to detect hand washing activity and the hand washing techniques as well. The classification results of getting the soap and rubbing thumbs are low, which will be the main focus in the future study.

RESUMEN

Monitoring activities of daily life (ADLs) allows to evaluate health conditions for older adults. However, there are still a limited number of studies on bathroom activities monitoring using a wrist-mounted accelerometer. To fill this gap, in this study, researchers collected data from 15 older adults wearing a wrist-mounted accelerometer. Six bathroom activities, i.e., dressing, undressing, brushing teeth, using toilet, washing face, and washing hands, were investigated. In total, 49.4-hour data for bathroom activities were collected. A hybrid convolutional neural network (CNN) is introduced for bathroom activity recognition. This hybrid CNN model is developed using both hand-crafted and CNN-based features as input. The proposed hybrid CNN model is compared to four machine learning models, i.e., Multilayer Perceptron (MLP), Support Vector Machines (SVM), K-nearest Neighbors (KNN), and Decision Trees (DT), and a conventional CNN model. Based on the classification results of leave-one-subject-out cross-validation (LOSO), the hybrid CNN model outperformed the other models. The hybrid CNN model is also tested based on a transfer learning method. As a calibration step based on LOSO, the transfer learning method additionally trains the model with an example of each activity from the test subject. The transfer learning method obtained better classification performance than LOSO. With transfer learning, the f1-score for using toilet was improved from 0.7784 to 0.8437. This study proposes a deep learning model fusing hand-crafted features and CNN-based features. Besides, the transfer learning method offers a way to build subject-dependent models to improve the classification performance.Clinical relevance -This provides a model that helps monitoring older adults' bathroom activities using a single wrist-mounted accelerometer.

Asunto(s)

RESUMEN

The purpose of computer-aided diagnosis (CAD) systems is to improve the detection of diseases in a shorter time and with reduced subjectivity. A robust system frequently requires a noise-free input signal. For CADs which use heart sounds, this problem is critical as heart sounds are often low amplitude and affected by some unavoidable sources of noise such as movement artifacts and physiological sounds. Removing noises by using denoising algorithms can be beneficial in improving the diagnostics accuracy of CADs. In this study, four denoising algorithms were investigated. Each algorithm has been carefully adapted to fit the requirements of the phonocardiograph signal. The effect of the denoising algorithms was objectively compared based on the improvement it introduces in the classification performance of the heart sound dataset. According to the findings, using denoising methods directly before classification decreased the algorithm's classification performance because a murmur was also treated as noise and suppressed by the denoising process. However, when denoising using Wiener estimation-based spectral subtraction was used as a preprocessing step to improve the segmentation algorithm, it increased the system's classification performance with a sensitivity of 96.0%, a specificity of 74.0%, and an overall score of 85.0%. As a result, to improve performance, denoising can be added as a preprocessing step into heart sound classifiers that are based on heart sound segmentation.

Asunto(s)

RESUMEN

This study introduces machine learning predictive models to predict the future values of the monitored vital signs of COVID-19 ICU patients. The main vital sign predictors include heart rate, respiration rate, and oxygen saturation. We investigated the performances of the developed predictive models by considering different approaches. The first predictive model was developed by considering the following vital signs: heart rate, blood pressure (systolic, diastolic and mean arterial, pulse pressure), respiration rate, and oxygen saturation. Similar to the first approach, the second model was developed using the same vital signs, but it was trained and tested based on a leave-one-subject-out approach. The third predictive model was developed by considering three vital signs: heart rate (HR), respiration rate (RR), and oxygen saturation (SpO2). The fourth model was a leave-one-subject-out model for the three vital signs. Finally, the fifth predictive model was developed based on the same three vital signs, but with a five-minute observation rate, in contrast with the aforementioned four models, where the observation rate was hourly to bi-hourly. For the five models, the predicted measurements were those of the three upcoming observations (on average, three hours ahead). Based on the obtained results, we observed that by limiting the number of vital sign predictors (i.e., three vital signs), the prediction performance was still acceptable, with the average mean absolute percentage error (MAPE) being 12%,5%, and 21.4% for heart rate, oxygen saturation, and respiration rate, respectively. Moreover, increasing the observation rate could enhance the prediction performance to be, on average, 8%,4.8%, and 17.8% for heart rate, oxygen saturation, and respiration rate, respectively. It is envisioned that such models could be integrated with monitoring systems that could, using a limited number of vital signs, predict the health conditions of COVID-19 ICU patients in real-time.

Asunto(s)

RESUMEN

BACKGROUND: Although deep neural networks (DNNs) are showing state of the art performance in clinical gait analysis, they are considered to be black-box algorithms. In other words, there is a lack of direct understanding of a DNN's ability to identify relevant features, hindering clinical acceptance. Interpretability methods have been developed to ameliorate this concern by providing a way to explain DNN predictions. METHODS: This paper proposes the use of an interpretability method to explain DNN decisions for classifying the movement that precedes freezing of gait (FOG), one of the most debilitating symptoms of Parkinson's disease (PD). The proposed two-stage pipeline consists of (1) a convolutional neural network (CNN) to model the reduction of movement present before a FOG episode, and (2) layer-wise relevance propagation (LRP) to visualize the underlying features that the CNN perceives as important to model the pathology. The CNN was trained with the sagittal plane kinematics from a motion capture dataset of fourteen PD patients with FOG. The robustness of the model predictions and learned features was further assessed on fourteen PD patients without FOG and fourteen age-matched healthy controls. RESULTS: The CNN proved highly accurate in modelling the movement that precedes FOG, with 86.8% of the strides being correctly identified. However, the CNN model was unable to model the movement for one of the seven patients that froze during the protocol. The LRP interpretability case study shows that (1) the kinematic features perceived as most relevant by the CNN are the reduced peak knee flexion and the fixed ankle dorsiflexion during the swing phase, (2) very little relevance for FOG is observed in the PD patients without FOG and the healthy control subjects, and (3) the poor predictive performance of one subject is attributed to the patient's unique and severely flexed gait signature. CONCLUSIONS: The proposed pipeline can aid clinicians in explaining DNN decisions in clinical gait analysis and aid machine learning practitioners in assessing the generalization of their models by ensuring that the predictions are based on meaningful kinematic features.

Asunto(s)

RESUMEN

Datasets in healthcare are plagued with incomplete information. Imputation is a common method to deal with missing data where the basic idea is to substitute some reasonable guess for each missing value and then continue with the analysis as if there were no missing data. However unbiased predictions based on imputed datasets can only be guaranteed when the missing mechanism is completely independent of the observed or missing data. Often, this promise is broken in healthcare dataset acquisition due to unintentional errors or response bias of the interviewees. We highlight this issue by studying extensively on an annual health survey dataset on infant mortality prediction and provide a systematic testing for such assumption. We identify such biased features using an empirical approach and show the impact of wrongful inclusion of these features on the predictive performance.Clinical relevance- We show that blind analysis along with plug and play imputation of healthcare data is a potential pitfall that clinicians and researchers want to avoid in finding important markers of disease.

Asunto(s)