Integrating Multi-sensor Time-series Data for ALSFRS-R Clinical Scale Predictions in an ALS Patient Case Study

Abstract

Clinical tools for tracking functional decline in amyotrophic lateral sclerosis (ALS) rely on in-clinic guided assessments, such as the gold standard ALS Functional Rating Scale Revised (ALSFRS-R) instrument, thus limiting the frequency of collection and potentially delaying needed treatments. As such, ALS clinicians may miss subtle yet critical shifts inpatient health -pointing to the needfor objective and continuous capturing of day-to-day functional status. In-home health sensors could supplement clinical instruments with more frequent, quantitative measurements as early indicators of change. Using the XGBoost regressor in base learning, we explore interpolation techniques for aligning monthly ALSFRS-R assessment targets with high frequency sensor-based health features. We evaluated 9 interpolation models, which demonstrate superior prediction of ALSFRS-R scores compared to traditional clinical scale estimates based on linear slope. This pilot work provides a practical approach of modeling mixed-frequency data and shows the potential of using sensor-based health estimates as sensitive prognostic markers.

Introduction

Prior research at the Center to Stream Healthcare In-Place (C2SHIP) [1] at the University of Missouri (MU) has developed ubiquitous sensor systems designed to support aging-in-place through home health monitoring, encompassing measurements for cardiovascular, respiratory, sleep quality, gait, and room-level activity [2, 3]. Longitudinal studies conducted through the MU Sinclair School of Nursing have demonstrated how continuous monitoring can positively impact the health of aging populations by facilitating timely interventions and personalized care approaches [4]. These findings affirm the value of integrating sensor-based monitoring systems for tracking functional measures and adverse event risk into routine care, especially for those at heightened risk of incidents due to physical vulnerabilities, home confinement, and social isolation [5].

Sensor technologies have application beyond geriatric care, showing promise for monitoring patients with conditions characterized by limited functional range [6] and stroke recovery [7]. There is emerging evidence supporting the use of sensor and wearable technologies for monitoring amyotrophic lateral sclerosis (ALS), demonstrating their efficacy in capturing the biophysical and functional indicators necessary for managing this condition [8, 9]. Currently, ALS is established and periodically measured for disease stage progression using the in-clinic ALS Functional Rating Scale Revised (ALSFRS-R) scale [10]. The ALSFRS-R is an instrument for tracking changes from an initial baseline evaluation across bulbar (speech, salivation, and swallowing), fine motor (handwriting, cutting and handling utensils, dressing and hygiene), gross motor (turning in bed, walking, climbing stairs), and respiration (dyspnea, orthopnea, respiratory insufficiency) functional scales. Disease progression is estimated using the linear slope of the composite ALSFRS-R score over time, reflecting functional decreases in motor and respiratory areas. However, non-linear trajectory patterns in clustered ALSFRS-R scores have been detected across a range of clinical trial study data, providing evidence that ALS disease progression can be marked by a period of stability followed by steep decline [11]. Additionally, other studies suggest that more accurate patient-specific prognosis can be achieved using the slope of individual ALSFRS-R component scores as a multi-dimensional measure in place of a single composite score [12].

Multi-Rate Data Fusion Challenge

The disparity in sampling rates between high-frequency sensor data and clinician guided assessment scales presents a unique challenge for building predictive models in the clinical setting with integrated multi-modal features, particularly in the context of tracking progressive diseases like ALS where assessments are conducted only periodically. Various methodological approaches have been explored for merging disparate temporal data. One approach is to aggregate the sensor data to match the periodicity of assessment instruments [13, 14]. However, aggregation of the time-series between assessment points lends to a loss of fidelity in the captured sensor data features and dramatically reduces the number of observations in the resultant feature space [15]. Alternatively, employing interpolation techniques to increase the temporal resolution of the sparse clinical scores to align with the sensor data may provide improved representation of patient status over time [16]. The use of temporal aggregation or interpolation to align wearable and ambient sensor data predictors and clinical scale label vectors for predictive machine learning has been applied widely across healthcare domains, including in neurodegenerative disease [17] and cardiovascular disease [18] assessment models. Moreover, advancements in neural network architectures and embeddings have introduced more sophisticated integration techniques. These include self-supervision data augmentation [19, 20], deep learning-based interpolation [21], and the application of auto-encoders for labeling high-frequency data [22]. While extending the use of machine learning to self-labeling of sensor data offers promising avenues for improving the accuracy of predictive models, for single participant case study data with limited feature space and infrequent clinical measures we focus on interpolation methods to facilitate the temporal sampling of clinical assessments for supervised learning.

Methods

Setting and Participants

Clinical ALSFRS-R scales and remote sensor-based health parameters are being collected as part of a single-site single-cohort prospective study evaluating around the clock monitoring of ALS patients with in-home sensor technology. The study has been approved and is overseen by the MU Institutional Review Board. Routine clinic visits are typically scheduled every 3 months and care teams may not be aware of decline in function between visits. The study aims to assess the use of remote sensor applications for monitoring ALS progression, particularly investigating whether observational sensor data are predictive of functional decline in ALS patients. Participants are screened and consented at the MU Health ALS Clinic, inclusion criteria are: ALS diagnosis, residing within 100 miles of the clinic, with either a home caregiver or a Montreal Cognitive Assessment (MoCA) score greater than 22. The ALSFRS-R scales, shown in Figure 1, were collected over a two year period between 2022 and 2024 from a single patient. Initial assessments were collected at quarterly clinic visits to month 9, and currently are collected every month. In months that a clinic visit has been scheduled, nursing staff at the MU Health ALS Clinic conducts an assessment over the telephone in the week prior to the scheduled visit and enters the scored instrument into the patient’s chart. An occupational therapist (OT) at the MU College of Health Sciences then retrieves the scores manually from the patient’s electronic medical record and transfers them to a research data repository for analysis. In months for which a clinic visit is not scheduled, the OT conducts a telephone assessment with the patient ensuring consistent monthly assessments.

Figure 1:

Collected ALSFRS-R scales over time arranged by respective functional area.

Data Description

Remote online sensor systems have been continuously collecting and transferring data from within the participant’s home since early 2023 to a secure processing server. A hydraulic bed mat positioned on the side of the bed that the participant sleeps on collects respiration, pulse, and sleep restless measurements. Motion sensors in the bathroom, bedroom, kitchen, living room, and above the front entrance, monitor room-level activity. Participant gait parameters were captured by a 3D thermal depth camera sensor monitoring the living room floor plane using a point-cloud stride detection algorithm. However, the participant transitioned to a wheelchair shortly after enrollment and gait parameters have not been collected for a substantial portion of the observation period. As such, models included in the analysis were trained using only the hydraulic bed mat and motion sensor health parameters. The raw sensor data are algorithmically processed to extract tabulated features prior to modeling. The bed mat sensors contains four hydraulic transducers with multiple pulse rate extraction algorithms applied to the ballistocardiogram signal for each transducer. Pulse rate is estimated using Hilbert transform [23], windowed peak to peak deviation [24], short-time peak energy [25], and k-means clustering [26]. Respiration rate is calculated using a count of zero-crossings in the signal [24]. Motion sensor activity is collected as 7-second epochs and aggregated to 15 minutes, hourly, night, day, and 24-hours and then processed for room density over time [27].

The ALSFRS-R scales for salivation, dypsnea, and respiratory were excluded from the modeling analysis due to having constant values with zero variance in their respective scores. Decreases in speech, swallowing, handwriting, cutting, and turning scores beginning on Month 13 suggests a rapid loss of function, as reflected in the composite score dropping eight points from 31 to 23, shown in Figure 2. Because model training and test datasets are split temporally to retain the sequential nature of the collected scores, rather than with random selection, training set observations do not present the same labels as in the test dataset. We include time series (serial date, hour of day) and 7-day lagged features in the modeling, commonly used techniques to forecast unseen time-dependent target labels, to account for changes in ALSFRS-R scales over time.

Figure 2:

Effect of interpolation techniques on the ALSFRS-R composite score.

Integrating ALSFRS-R Scales and Sensor Data

Values for missing sensor features on days when observations were available for other sensor types were left intact with the expectation that missing data may be representative of participant state over the reference time period. For example, when the participant doesn’t sleep in their bed or enter the bathroom over a period of time the null values are due to changes in behavior or function. This is also the rationale for the use of extreme gradient boosting model (XGBoost) [28] for native handling of null values.

We evaluated integrating the individual component ALSFRS-R scores and composite score labels with the daily aggregated sensor health features using six possible interpolation techniques. The lower frequency ALSFRS-R scores were sorted and date indexed adding a row for every date of collected sensor data, filling the dates between, using the beginning and end dates of the sensor health features. Linear interpolation was applied in two passes, forward and then backward. The linearly interpolated ALSFRS-R scales serve as the baseline as a linear slope in functional decrease is expected in ALS patients by clinicians. Backward fill interpolation was applied to associate prior sensor measurements with future collected scores. Cubic interpolation was conducted using a third order spline to produce a smoothed curve of gradual increase and decrease in function with peaks located at collection points. Exponential and inverse exponential were done using a scale factor of 10 to simulate the cases for disease progression changing nearest to the prior collected score or to the current score, respectively. Sigmoid interpolation was applied with beta value of 0, which combined the shapes of the two exponential interpolations to form an a sinusoidal curve with positive or negative peaks for cases where an increase is followed by decrease or decrease followed by increase, effectively smoothing the line and producing a peak in the waveform. Examples of each temporal interpolation technique are illustrated in Figure 2.

Predictive Models

Summary statistic features were engineered for each sensor data parameter in Table 1, using the parameter mean, median, mode, minimum, maximum, sum, standard deviation, count, variance, range, first value, and last value. The summary statistics were calculated within each 24-hour period in the 30-days leading to the date of ALSFRS-R assessments. The generation of summary features from the raw derived sensor measurements increased the number of dataset dimensions, expanding from 128 to 2361 features. Features with high colinear correlation were removed by iterating through the correlation matrix columns and for each pair of correlations greater than a cutoff of 0.9 the feature with the larger average correlation was dropped. The training feature set was selected as the earliest 0.8*number of dates temporally in the dataset, while the test hold out set was selected as the remaining 0.2*number of dates.

Table 1:

Sensor data parameters.

Sensor Parameter	Description
Pulse rate min	Pulse rate measures deconvolved from ballistocardiogram signal across four hydraulic transducers using hilbert transform, windowed peak to peak deviation, short-time energy, and k-means clustering.
Pulse rate mean
Pulse rate max
Pulse rate confidence
Pulse rate transducer
Respiration rate min	Respiration rate measures deconvolved from ballistocardiogram signal across four hydraulic transducers using zero crossing count.
Respiration rate max
Respiration rate mean
Respiration rate confidence
Motion artifacts max	Motion signal peaks above 3 standard deviation threshold in ballistocardiogram signals across four hydraulic transducers.
Motion artifacts level
Motion artifacts transducer
Raw transducer	Raw ballistocardiogram signals across four hydraulic transducers.
Raw transducer sum
Duration restless	Duration measures derived from signal onsets and offsets detected in ballistocardiogram signal across four hydraulic transducers.
Duration restless max
Duration in bed
Motion count	Motion sensor activity for bathroom, bedroom, front door, kitchen, living room
Motion density
Duration bathroom visit	Duration measures derived from motion counts within bathroom and residence, further aggregated to day and night
Duration vacant

An initial model was built with all features, hyperparameter tuned using a grid search as in Table 2, and evaluated on the test hold out data using 5-fold cross validation for each interpolation type. To prevent model overfitting we included regularizing hyperparameters for ‘min_child_weight’, ‘lambda’, and ‘alpha’. Feature importance weights of each feature across all splits in boosted trees were then extracted from the initial model and binned by thresholding the feature importance scores to the nearest 0.001. Each set of thresholded features were subsequently fitted with a non-trained XGBoost regressor estimator, hyperparameter tuned, and evaluated on the test holdout set using 5-fold cross validation. The best model for each interpolation method and ALSFRS-R sub-scale pairing were then saved and the resulting test errors were compared across all models.

Table 2:

Hyperparameter tuning search space.

Parameter	Values	Parameter	Values
max_depth	3, 5, 7, 9, 12, 15, 17, 25	subsample	0.6, 0.7, 0.8, 0.9, 1
learning_rate	0.001, 0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5	colsample_bytree	0.6, 0.7, 0.8, 0.9
min_child_weight	0.1, 0.5, 1, 2, 3, 4, 5, 6, 7	n_estimators	128, 384, 640, 896, 1152
lambda	0.01, 0.1, 0.3, 0.5, 1, 3, 5, 10, 25, 50, 100, 150, 200	alpha	0, 0.01, 0.1, 0.3, 0.5, 1, 3, 5, 10, 25, 50, 100, 150, 200
gamma	0	objective	reg:squarederror

Results

We compare initial models fit on all features against models fit with feature selection for each interpolation type and models fit with linear interpolation using root mean square error (RMSE) and mean absolute error (MAE) as the primary metrics of model performance. This section outlines the specifics of the best model for each ALSFRS-R component scale based on the interpolation method and feature selection in model optimization as referenced in Table 3. The composite model had the highest error with an initial RMSE of 3.391 and MAE 2.764 and best fit RMSE of 2.984 and MAE 2.497 using 77 features and backfill interpolation, which outperformed the linear model RMSE 3.883 and MAE 3.216.

Table 3:

Model performance metrics.

Bulbar Functions

Speech sub-scores were best fit with a model using backward fill interpolation predicted by thirteen features. There was a modest improvement in error between the initial model and the model with feature selection, from initial RMSE 0.814 to 0.788 and MAE from 0.659 to 0.638, and compared to the linear model RMSE 1.153 and MAE 0.973. Beginning at month 13 we see a decrease in speech from 4 to 2, aligning with the test train split date, which may lend to the poor model fit. The prediction model for swallowing improved using exponential interpolation, which suggests that the trends in previous sensor data corresponded with more gradual changes in the swallowing scores. Swallowing scores had decreased in month 9 from 4 to 3 before returning to 4 in month 10, which helped the model to more accurately predict on the test data after month 14 where the score decreased again from 4 to 2. The initial model RMSE of 0.447 reduced to 0.396 with feature selection, and MAE from 0.243 to 0.221, a slight improvement from the linear model RMSE 0.521 and MAE 0.309. The feature count was optimized at 21.

Fine Motor Functions

Improvements in handwriting sub-score prediction between the initial model and the feature selection model in RMSE from 0.496 to 0.467 and MAE from 0.317 to 0.292 was observed compared to the linear model RMSE 0.572 and MAE 0.39. The final model utilized 33 features and exponential interpolation of labels, indicating a slightly more complex relationship amongst predictors for handwriting due to the reliance on bed sensor features without gait information available. The difference between scores could be attributed to the removal of two lagged features, ‘date_lag_4’ and ‘date_lag_6’. For the cutting sub-score models, backfill interpolation is more predictive. The over time trend in cutting was similar to speech, which also selected for backfill interpolation, where the scores followed a constant non-changing period followed by decline after the date of the test train split. The RMSE improved slightly from 0.182 to 0.11, and MAE from 0.168 to 0.105 an improvement from the linear model RMSE 0.65 and MAE 0.566, with 11 features in the final model. Dressing exhibited the highest amount of variability in the collected scores, with notable decreases followed by increases in months 9 to 12 and months 13 to 15. The model using sigmoid interpolation for dressing sub-scores performed best as the function pushed the curve at the inflection points together, in effect forming an outlier peak in the score which decreased the model error for the leading and following dates within the troughs. The dressing model RMSE decreased from 0.263 to 0.126, and MAE from 0.203 to 0.079 in compared to the linear model RMSE 0.415 and MAE 0.322. The best model selected three time series features of near equivalent importance weight handwriting_lag_1 weighted at 349.0, dressing_lag_1 at 349.0, and dressing_lag_2 at 289.0.

Gross Motor Functions

The model for turning utilized 10 features, seven bed signal features and and three time series features including serial date and lagged features for cutting and dressing, suggesting that changes in gross motor ability coincide somewhat with changes in fine motor function. The initial RMSE improvement from 0.147 to 0.029 and MAE from 0.13 to 0.023, decreased from the linear model RMSE 0.677 and MAE 0.629. The model relied on backfill interpolation reflecting that prior sensor data was predictive of changes the change in ability to turn and adjust in bed. The collected scores for walking followed a similar trend with turning, however the model did not perform as well and required additional features increasing the selected feature count to 31. RMSE decreased only slightly from 0.596 to 0.571 and MAE from 0.421 to 0.405. The linear model slightly outperformed the best model with an 0.558 RMSE and 0.398 MAE, the only linear model to do so. The best model used the cubic spline interpolation, which is most similar to the linear trend in shape. Stairs had near perfect accuracy for the initial, best, and linear model. This can be attributed to the decrease in the scores well before the test train split, in month 7. However the model relied on only three time series features weighted equally at 1 for serial_date, swallowing_lag_7, and ‘stairs_lag_1’, using inverse exponential interpolated scores and without sensor data features.

Respiratory

The orthopnea model likewise showed a perfect accuracy for RMSE and MAE, with slight improvement from the initial model RMSE of 0.088 and MAE of 0.084 and linear model RMSE 0.035 and MAE 0.001 using an exponential interpolation curve. The model utilized six time series features, four ‘orthopnea_lag’, ‘dressing_lag_6’, and ‘swallowing_lag_2’. Given the decrease and then increase in scores between months 7 and 12 from 4 to 2 to 4, there is a possibility that changes in treatment that might be extracted from EHR as predictors.

Feature Selection

We also evaluated feature sets for the ALSFRS-R scales by aggregating the features by type over the score interpolation models to determine which sets of predictors were most utilized by XGBoost according to the weighted frequency of the splits in boosted trees, show in Figure 4. We see that the raw bed sensor data was predominant for all models except dressing and orthopnea. The 7-day lag variables were more predictive for dressing and orthopnea. The derived bed algorithm features, time of day variables, and serialize date were only slightly important. The remaining features for sleep restlessness, bathroom visits, motion density, residence vacancy, and time in bed were rarely used for splits.

Figure 4:

Feature selection counts by ALSFRS-R sub-scale models summated across interpolation techniques.

Discussion

The objective for this paper was to establish a framework for integrating and analyzing sensor-derived data with clinical assessments to enhance the granularity and frequency of functional assessment in ALS patients. With precision modeling in mind, we investigated the use of target label interpolation techniques for improving prediction error of functional states over baseline linear change in slope. The variability in the number of features selected across models, from as few as 3 in the case of Dressing and Stairs to as many as 33 for Handwriting and 77 for the total composite, highlights the potential of precision models in capturing underlying ALS symptom deterioration from sensor health features. Fine and gross motor area models for Dressing, Cutting, Turning, and Stairs, which were fit with a minimal number of features selected relied primarily on the time lag features, suggesting that absent of motor-related sensor data the temporal progression in functional decline in ALS plays an important role in predicting future changes. However, conversely, the increase in number of features required for accurately modeling Speech (13) and Swallowing (21) indicates that the individual sensor data features, such as pulse rate and respiration, may not best capture changes in component ALSFRS-R scales over time and additional sensor data types and motor function measures are needed. As additional participants are recruited, we are especially interested in capturing gait from early stage patients as the gait measures are expected to correlate with gross and fine motor functional scales.

Data Stream Modeling

For implementing predictive models in a real-world scenario, for instance as part of a clinical decision support system (CDSS), it would be useful to tailor the interpolation application to the unique pattern of functional decline progression determined from trends in existing ALSFRS-R measurements. Due to the homogeneous nature of ALS disease and variation in home environments and lifestyles, it is expected that future sensor-based assessment algorithms will rely on personalized models trained and evaluated on individual participant sensor data. Particularly, the decrease in scores over time leads to the possibility of unseen labels in the test data which do not appear in the training data. For this reason, it would be ideal to use a sequential learning algorithm with a method for updating the model with new data input, such as Long Short-Term Memory (LSTM) networks. Sequential learning models could be incorporated into the data processing pipeline after derived features are calculated with the daily predictions stored and retrieved through a clinical support interface.

Study Limitations and Future Work

One significant limitation is the analysis of single-participant data, which doesn’t capture the variability in symptom progression and response to treatment observed across the broader ALS population and as such, the findings will not generalize as well as group level models. The absence of gait information further narrows the scope of the results. Given the role of ambulatory function in assessing ALS progression, having sensor-based motor control measures would be beneficial for the gross and fine motor function ALSFRS-R sub-scale models. Our approach also did not incorporate machine learning self-supervision methods for data labeling, which could potentially enhance the accuracy and reliability of symptom tracking by refining the dataset or deep learning models for self-supervision labeling, which could potentially increase the accuracy for predicting functional decline progression. This limitation underscores the potential for future research to employ more advanced machine learning strategies and group level data for further understanding of ALS disease progression.

Conclusion

In this study, we identified the model parameters for generating predictive analytics that support sensor-based in-home monitoring of amyotrophic lateral sclerosis disease progression from single participant case-study trial data. We employed machine learning using the XGBoost regressor estimator to predict the progression of amyotrophic lateral sclerosis as measured by the ALSFRS-R clinical instrument, focusing on an individual patient’s component ALSFRS-R scales related to Speech, Swallowing, Handwriting, Cutting with utensils, Dressing, Turning in bed, Walking, Climbing stairs, and Orthopnea. Salivation, Dyspnea, and Respiratory ALSFRS-R scores for this individual were not evaluated due to a lack of variance in those scores. Our evaluation of hyperparameter optimization and feature selection revealed notable improvements in model accuracy across the component scales and the summated composite. The results demonstrate the potential of sensor-based assessments to predict and track functional decline progression, offering insights into the disease’s progression outside of the clinical setting that could lead to more timely and personalized care interventions. In our analysis of applying multiple interpolation techniques and iterative feature selection-learner, we have shown that trends in individual scores over time correspond to differing interpolation shapes, suggesting that non-linear approaches to estimating disease progression within component scale ranges are appropriate. As such, predictive models should be fit for individual ALSFRS-R components with varying feature selection and interpolation for best accuracy. Utilizing ALSFRS-R predictions could significantly enhance the timing and precision of interventions, ultimately improving the quality of life for individuals affected by ALS.

Figures & Tables

Figure 3:

Model processing pipeline.