Introduction

Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease marked by degeneration of motor neurons, leading to eventual muscle paralysis. Globally, ALS has an incidence rate of 1–2 per 100,000 individuals1, with a lifetime risk of approximately 1 in 3502. As the disease progresses, patients experience continued decline in motor function, albeit at variable rates. The complex nature of ALS and the heterogeneity of its presentation, progression, and rate of decline pose significant challenges for timely diagnosis and statistical analyses in ALS trials3,4,5.

The Revised ALS Functional Rating Scale (ALSFRS-R) is the most frequently used functional assessment in ALS trials, despite its shortcomings, including its reliance on subjective perception and insensitivity to small changes in fine motor function6,7. There is a pressing need for more sophisticated, objective, and accessible metrics to assess and monitor ALS.

Because loss of fine motor function is common in people with ALS (PALS) due to hand weakness, stiffness and/or slowness, quantifying fine motor impairment by analyzing keystroke dynamics during smartphone use presents a promising avenue to quantify motor changes due to ALS. Keystroke dynamics, the quantitative metrics of keyboard interactions, including pressure, timing and rhythm of an individual’s typing patterns, have been shown to reflect changes in fine motor function in several neurological diseases, including Alzheimer’s, Parkinson’s, Huntington’s and multiple sclerosis8,9,10,11,12. A prior study examined typing activity in ALS13, but recent technological advancements allow for more refined insights. Keystroke dynamics monitoring offers high resolution and real-time data that can be accessed remotely. To exploit the advantages of this monitoring technique, we developed a method integrating acquired personal smartphone keystroke data with machine learning analytic techniques to harness the full potential of keystroke dynamics to detect and monitor ALS progression (nQiALS toolkit, pronounced “n-q-i-a-l-s”).

The purpose of this study was to explore the feasibility and utility of keystroke dynamics in detecting, quantifying, and monitoring fine motor impairment and progression in people with ALS by (1) investigating whether the nQiALS toolkit can identify characteristic typing patterns in PALS and (2) evaluating the utility of the nQ-derived metrics (from nQ Medical, currently managed by Area2 AI, Cambridge, MA) as digital biomarkers to monitor fine motor decline over time in people with ALS. We present the design, development, implementation, and results of the nQiALS toolkit.

Results

Participant characteristics and adherence

In this single-center, observational study, 93 participants were enrolled, comprising both those with a clinical diagnosis of ALS and age- and sex-matched healthy controls (HC) (Table 1). Participants underwent an initial screening/baseline visit (V1), followed by three follow-up visits at intervals of 3, 6, and 9 months (V2-4).

Table 1 Participant demographic and clinical characteristics.
Full size table

No study visits were missed. Out of the 93 participants, there were 11 cases of early termination (8 ALS, 3 HC), 2 lost to follow-up (1 ALS, 1 HC), and one participant with ALS passed away due todisease progression. Reasons for early termination included loss of arm/hand mobility (n = 2 PALS), overall rapid progression of the disease (n = 3 PALS), cognitive decline (n = 1 PALS), displeasure with the keyboard (n = 1 PALS), and scheduling difficulties (n = 1 HC). Additionally, three participants dropped out prematurely for unknown reasons (1 PALS, 2 HC).

Raw typing and keystroke features

At V1, participants installed the nQ keyboard software on their smartphones and were instructed to use it as they typically would. The software recorded keystroke dynamics unobtrusively over the follow-up period as participants used the nQ keyboard. The keyboard had to be selected in place of the default keyboard and remained as the new default keyboard until switched. A typing session began automatically with any new keyboard input and concluded upon text input halt or app switching. During the study a total of 1.06 GB of remote keystroke data were collected, comprising 1.2 million sessions recorded across all participants. The mean percentage of active days (i.e., days with at least 5 keystroke sessions recorded) was 54% of days overall, with no significant differences between PALS and HC. The average number of keystrokes per session was 40 (SD 15), with no significant differences between PALS and HC. Table 2 presents a summary comparison of the typing data collected in the two groups, with typing metrics evaluated over the full data collection period.

Table 2 Typing characteristics and keystroke metrics.
Full size table

While total number of keystrokes did not differ significantly between the two groups, PALS had significantly longer typing sessions (p < 0.001). In the natural typing analysis, PALS typed slower reflecting the overall longer flight times (time between consecutive key presses), slower velocity, and longer hold times compared to the HC group.

Disease detection

The nQiALS-Detection model was developed to detect the likelihood a participant had ALS. The model optimized a support vector machine (SVM) classifier—a machine learning technique renowned for its effectiveness in classification tasks, even with limited data—to differentiate keystroke patterns between PALS and HC. The nQiALS-Detection scores from PALS differed from HC significantly with an AUC of 0.89 (p < 0.001). The median [IQR] detection score for the ALS group was 0.35 [0.26–0.37], while for the HC group it was 0.20 [0.2–0.22], with a p value of < 0.001. When comparing PALS with mild symptoms (“Mild ALS,” defined here as ≥ 9 in the ALSFRS-R fine motor subdomain) versus HC, the AUC was 0.93 (p < 0.01) (Fig. 1). In comparison, the alternating finger tapping (AFT) test showed a lower AUC of 0.67 (p < 0.05) for PALS vs HC, and an AUC of 0.52 for Mild ALS vs HC (NS). The results showed that nQiALS-Detection scores were not significantly impacted by sex or age, supporting the model’s consistency across various demographic groups.

Figure 1
figure 1

Comparing the performance of the nQiALS-Detection model with Fingertapping. The nQiALS-Detection score represents the likelihood that the input typing data originates from an ALS sample, with higher scores indicating a stronger likelihood of ALS. (a) The model’s performance was tested based on its ability to separate ALS samples from HC, quantified using the receiver operating characteristic area under the curve (ROC AUC) for HC versus all people with ALS. (b) As a surrogate for early detection, performance was also evaluated using the ROC AUC built on the separation between HC and mild ALS. ALS—all participants with ALS, mild ALS—participants with ALS who scored ≥ 9 points in the ALSFRS-R fine motor subdomain (questions 4–7) at baseline, HC—healthy control participants.

Full size image

nQiALS-Detection scores negatively correlated with the total ALSFRS-R and with the fine motor subdomain scores (Spearman ρ of − 0.49 (p < 0.001) and − 0.44 (p < 0.001), respectively). AFT had similar correlations with total ALSFRS-R scores and the fine motor subdomain (0.42 (p < 0.001) and 0.49 (p < 0.001), respectively).

Predicting ALS progression

The nQiALS-Progression model was developed to estimate overall ALS progression rate (δALSFRS-R), as measured by changes in total ALSFRS-R score, using a longitudinal stream of natural typing data. This model leverages recurrent neural networks (RNN) to predict the progression rate based on longitudinally collected keystroke data. The model output was then correlated against δALSFRS-R.

The correlation of nQiALS-Progression with the ALSFRS-R total score increased for approximately 12 weeks as the model gained data with which to predict the ALSFRS-R. After week 17, the correlation decreased (Fig. 2a). Overall, the correlations exhibited low-to-moderate strength, taking values of around 0.2, and were not significant at any time point.

Figure 2
figure 2

Evaluating ALS disease decline rate using nQiALS-Progression model. (a) Weekly Spearman ρ correlation between the weekly nQiALS-Progression and the overall δALSFRS-R over the 9-month follow-up period, and (b) AUC discriminating between slow and fast progressors for different δALSFRS-R thresholds (− 0.10, − 0.20, − 0.30, and − 0.40 points per week).

Full size image

The model was also used to stratify progression rate using different thresholds for definition of “fast” and “slow” progressors based on the rate of decline of the ALSFRS-R. The model performed well at more negative thresholds, i.e., when the definition of “fast progression” was more stringent (δALSFRS-R threshold < − 0.3 points/week), distinguishing markedly faster progressors with a high AUC. Conversely, for less stringent thresholds, the AUC scores generally decreased over time, as the model had lower accuracy distinguishing “fast” from “slow” progressors using less extreme definitions (Fig. 2b).

The model's predictions were in limited agreement with the observed ALSFRS-R score (Fig. 3). However, the positive mean of differences suggested that the model's predictions tended to overestimate the rate of decline by 0.1 point/week on the ALSFRS-R, predicting more rapid progression than the observed ALSFRS-R slope (Fig. 3).

Figure 3
figure 3

Comparison of nQiALS-Progression predicted rate of decline to the observed rate of decline on the ALSFRS-R. Bland–Altman comparing the 9-month δALSFRS-R and the average weekly nQiALS-Progression estimate. The positive mean of differences suggested that the model's estimates tended to overestimate the rate of decline by 0.1 point/week on the ALSFRS-R, suggesting more rapid progression than the observed ALSFRS-R slope.

Full size image

Fine motor changes monitoring

Raw variation of keystroke features was analyzed to assess fine motor change predictive ability. Analysis was conducted using baseline data against the final week’s data. Only ΔError X comparison between ALS Change and HC groups reached significance via the Mann–Whitney U Test (Table 3).

Table 3 Changes in keyboard metrics from baseline to end of study.
Full size table

The nQiALS-Fine Motor model was developed to analyze data from multiple keystroke features to compare longitudinal typing patterns of participants with declining fine motor impairment over time to those with stable fine motor function as defined by the score on the fine motor subdomain of the ALSFRS-R. This model leveraged the RNN-based machine learning model to identify keystroke features showing change over time and those that were stable.

To evaluate this model, we defined three subgroups, (1) participants with ALS who had change on the ALSFRS-R fine motor subdomain (ALS change), (2) participants with no change on the ALSFRS-R fine motor subdomain (ALS no change), and (3) HC. The predicted nQiALS-Fine Motor scores, which represent the probability of change in fine motor function over time, were highest for the ALS change group, followed by the ALS no change group, and lastly the HC group. Scores for the ALS change group were significantly higher than the HC group at week 12, 24, and 36 (p < 0.001 at all timepoints). Scores for the ALS no change group were also significantly higher than the HC group at these timepoints (p = 0.044, 0.048, 0.014, respectively).

The model suggested changes in fine motor impairment beyond what was captured by the fine motor subdomain of the ALSFRS-R. The nQiALS-Fine Motor score difference between HC and ALS no change groups increased over time, possibly indicating detection of fine motor symptom progression in the ALS no change group.

Conversely, differences in nQiALS-Fine Motor scores between the ALS no change and ALS change groups decreased over time. This trend suggests that some participants initially classified as stable (no change group) may have transitioned to the ALS change group. This hypothesis was confirmed by applying an ablation test, a method used in machine learning to evaluate the importance of specific input characteristics to the model’s performance. This is done by removing (‘ablating’) certain features and observing the effect on the model’s output. In our case, the ablation test involved using the same approach on time-shuffled longitudinal signals, which effectively removes the temporal reference from the dataset. The results of the ablation test showed stable classification performance over the follow-up period, suggesting that our model could be predicting transitions from the no change group to the change group as sequences approached the end of the follow-up. While this interpretation is based on the classification results, further investigation is required to confirm the model’s ability to capture early changes not yet detected by the ALSFRS-R.

Figure 4 presents the predicted probabilities of change and corresponding AUC values over time for each group. Variations in the nQiALS-Fine Motor model output underscore its potential to discern subtle dimensions of fine motor change, which may not be fully captured by the traditional ALSFRS-R scoring method. A detailed comparison between nQiALS-Fine Motor predictions and changes in both the overall and fine motor-specific ALSFRS-R scores across the observation period can be found in Table 4.

Figure 4
figure 4

Performance of the nQiALS-Fine Motor model. (a) Predicted score indicating relative likelihood of functional change over time. (b) Area under the curve (AUC) illustrating discrimination between groups over the corresponding time period.

Full size image
Table 4 Comparison between nQiALS-Fine Motor predictions and changes in both the overall and fine motor-specific ALSFRS-R scores at different observation intervals.
Full size table

Discussion

This study demonstrates the feasibility and potential utility for using keystroke dynamics to aid in detecting, quantifying, and monitoring fine motor impairment and progression in PALS, and possibly finding use as an ALS clinical trial outcome measure.

This is the first real-world study and remote study of smartphone keystroke dynamics in ALS. It highlights the feasibility of collecting passive, frequent, longitudinal data outside the clinic. Building on prior research on specific keystroke features, our study harnessed machine learning models with multiple features for detection and prediction. By leveraging the power of advanced pattern recognition computational methods in concert with the richness of smartphone keystroke data, the nQiALS toolkit distinguished PALS from healthy controls, even in the case of mild fine motor dysfunction, and quantified change over time. The nQiALS-Progression and nQiALS-Fine Motor models identified functional changes over time, concordant with the ALSFRS-R, but with objective data about fine motor function. We provide preliminary evidence to suggest that the nQiALS-Fine Motor model may be able to detect early changes not yet captured by the ALSFRS-R, as demonstrated by the convergence of differences between the ALS change and ALS no change groups.

At the same time, certain limitations of our study must be acknowledged. The cohort size and semi-controlled nature of our nQiALS-Detection model might not capture the full spectrum of real-world clinical scenarios. External factors such as smartphone model and screen size, and internal factors, such as the use of auto-correct and predictive text, might influence the results, highlighting the need to expand the cohort and address these variables in future research. A significant challenge during the training and evaluation of this regressor-activated network was the limited variance in the ALSFRS-R slopes across our sample. Our ALS cohort exhibited relatively slow progression, which complicated the models’ ability to evaluate and establish a solid correlation and calls for validation in a more rapidly progressive cohort of PALS. This limitation underscores that while the nQiALS models can identify trends and potential changes over time, they do not predict future ALSFRS-R scores per se, but rather estimate changes based on longitudinal data correlated with the final ground truth at the end of the follow-up period. Furthermore, there were some data acquisition challenges, such as variability in typing frequency and consistency, which may have affected the robustness of the correlations observed. Addressing these challenges in future studies with a more diverse and rapidly progressing cohort will be crucial to enhancing and further validating the predictive capability of the nQiALS models. Furthermore, there were some data acquisition challenges. While the nQ keyboard mimics native smartphone keyboards, participants noted issues like autocorrection errors and lag times. This may have reduced the number of completed typing sessions in the study, though it did not dissuade users from continuing in the study. Our study ensured participants’ privacy by collecting only non-sensitive keystroke metadata (press and release times, key zones, and precision X and Y) without recording text content, specific keys pressed, or tap locations. This approach preserves communication privacy. Additionally, potential shared phone use could introduce occasional variability, but the requirement for participants to use their own devices and continuous data collection over weeks help mitigate this impact. Moreover, the nQiALS-Detection model was not tested against other motor-impairing conditions, limiting its specificity to ALS. Future studies should include ALS-mimics to validate the model’s clinical utility in distinguishing ALS from other similar conditions. This study was also not able to distinguish the phenotypic determinant of keystroke changes in PALS. Smartphone typing requires diverse motor and cognitive skills. Given the cognitive impairments in some PALS, future keystroke studies could consider including cognitive assessments such as the Edinburgh cognitive and behavioural ALS screen (ECAS)14. Furthermore, level of education, mood, and fatigue could affect keystroke data and may be factored in future studies. We do not expect that these factors would bias our results, but they could introduce variability and reduced discriminability. Finally, studies on independent and more diverse populations will be required to assess the generalizability of our findings.

There are numerous potential applications of the nQiALS toolkit, the clearest of which are as an aid for ALS diagnosis and as an ALS trial outcome to evaluate the effectiveness of experimental interventions. The toolkit has the potential to help characterize subgroups of patients with unique disease trajectories, and assist with clinical recommendations, such as provision of augmentative communication devices.

In short, this study highlights the potential of keystroke dynamics as a digital tool for detecting and monitoring fine motor impairment in PALS. This input could serve as a sensitive marker for fine motor function in ALS, even with mild symptoms. For its broader clinical adoption, standardization and reproducibility are essential. Keystroke dynamics might then complement traditional clinical assessments, offering more frequent monitoring in real-world settings.

Methods

Study ethics

The Institutional Review Board (IRB) at Massachusetts General Brigham (MGB) and the Information Security Office at Massachusetts General Hospital (MGH) approved the study. All data collection and management adhered to MGB, state, and national guidelines and regulations.

Recruitment

Participants with ALS were recruited from the multidisciplinary ALS clinic at MGH as well as through study recruitment materials posted on-site and sent via email blasts to an opt-in email distribution list through the Sean M. Healey & AMG Center for ALS at MGH. Individuals who had participated in previous observational studies who had agreed to be contacted regarding future research opportunities were also sent recruitment materials. Healthy controls were also recruited through an online MGB platform designed to support recruitment efforts in which individuals can search for actively enrolling studies and then click a link to request additional information. All participants were over 18, capable of providing informed consent, and could independently operate a smartphone. Participants were required to own and use a personal smartphone for at least 15 min per day (self-report). ALS participants met the El Escorial Criteria for possible, probable, lab-supported probable, or definite ALS. Non-neurological controls could not have a neurodegenerative or neurological diagnosis affecting typing performance and could not be first-degree relatives of individuals with known genetic ALS. Exclusion criteria included unstable medical or neurological disorders interfering with participation and unwillingness to use digital monitoring tools.

Study design

This was a single-center, observational study conducted from 10/30/2019 to 12/23/2022 at Massachusetts General Hospital (MGH), Boston, USA. It involved participants with a clinical ALS diagnosis (n = 63) and age- and sex-matched healthy controls (HC, n = 30). The study consisted of a baseline visit (V1) and three follow-ups at 3, 6, and 9 months (V2–V4) (Fig. 5). During V1, participants provided written informed consent and shared their medical and ALS history. Each visit included ALSFRS-R collection, medication review, and feedback on the keyboard application. Due to the COVID-19 pandemic, all study procedures had to be conducted remotely. As a result, additional clinical outcomes initially planned, including the ALS cognitive and behavioral scale (ALS-CBS), handheld dynamometry (HHD), and vital capacity (VC), were not collected.

Figure 5
figure 5

Schema of study design. The study consisted of 4 study visits, collected at baseline (V1) and at three-month intervals. Each study visit included a brief typing task using the nQ keyboard and collection of the ALS functional rating scale-revised (ALSFRS-R) as a clinical outcome measure for the ALS population. During the entire length of the study, keystroke dynamics of natural typing were continuously collected using the nQ keyboard.

Full size image

Typing data acquisition and keystroke features

At V1, participants installed the nQ keyboard software on their smartphone devices (iOS or Android), received an nQ study ID, and were instructed to use their phones normally, substituting use of the nQ keyboard for their native smartphone keyboard as frequently as possible. Participants reported their dominant hand and typical typing style. The nQ software continuously collected unsupervised keystroke data throughout the study duration without recording the text content or specific keys pressed. Specifically, press and release times, key zones (defined by six uniform vertical sections of the keyboard surface), and precision X and Y (relative distance between the tap and the center of the target key measured in pixels) were recorded.

Active data was collected at baseline and each follow-up visit (3, 6, 9 months). Participants were asked to perform: (1) a 5-min controlled copy task, which served as a standard typing sample, consisting of transcribing a standard text using the touchscreen keyboard on their smartphone, and (2) a 30-s alternating finger tapping task on their smartphone device, using each hand.

The main features constructed from the raw keystroke data included hold time, flight time, error X/Y, finger velocity, and acceleration. Hold time represents the time required for pressing and releasing a single key. Flight time is the latency between releasing a key and pressing the following one. Error X/Y for each keystroke represents the distance between the center of the finger tap on the screen and the center of the target key’s area within the keyboard. The distance is measured in both the X and Y axes, reflecting tap precision. Velocity refers to the speed of the finger moving from one key to another, calculated as the zone distance between two consecutively pressed keys divided by the flight time. Acceleration is the rate of change in velocity as the finger moves between keys, which can provide insight into the smoothness and control of the typing motion.

A day with at least five nQ keyboard sessions was deemed an “active day”. A session began when the nQ keyboard is triggered on screen for text input and ended when the keyboard was hidden, either post-input, screen lock, or due to application switching. The percentage of active days was calculated as the number of active days/time span in days using nQ app × 100.

Study staff monitored keystroke data weekly. If data gaps appeared, staff reached out to participants for troubleshooting. The nQ software was updated in May 2022 for enhanced responsiveness; 10 participants opted for this update.

Demographic and keystroke feature statistical analysis

We analyzed demographics, ALS characteristics, typing style, activity, and keystroke features. We applied the Mann–Whitney U test to discern statistical differences in typing data between groups. This non-parametric test assessed differences in typing characteristics, highlighting unique patterns. We also calculated the receiver operator curve (ROC) area under the curve (AUC) for each keystroke feature to gauge their ability to distinguish ALS from HC. Using AUC as a metric, we measured the classification capability of each keystroke feature before applying more advanced modeling techniques.

nQiALS-toolkit construction and performance evaluation

We developed the nQiALS toolkit, a set of three machine learning models that analyzed typing data and identified patterns related to motor performance in PALS. The three models include: (1) nQiALS-Detection, a classifier model that predicts whether a participant has ALS or not, (2) nQiALS-Progression, a progression model that predicts the weekly slope of decline for the total ALSFRS-R score, and (3) nQiALS-Fine Motor, a classifier model that predicts the weekly likelihood that the ALSFRS-R Fine Motor Subdomain score has declined. The nQiALS-Fine Motor model was built as a classifier, rather than as a regressor because there was little change in the ALSFRS-R Fine Motor Subdomain, making it difficult to train a progression model for this purpose. Python 3.12 was used for data analysis, model development and evaluation.

nQiALS-Detection

The nQiALS-Detection model was developed to detect the likelihood a participant had ALS. It was used to analyze a cross-sectional snapshot of typing patterns from PALS compared to matched HC. The input consisted of a series of descriptive statistics computed on the distribution of normalized flight times collected during the 5-min semi-controlled typing stream during V1. The target was a binary label as ground truth indicating whether a sample belonged to a HC (0) or PALS (1). The model was trained using sixfold cross-validation to ensure independence between training and testing sets. The final architecture employed a SVM Classifier with a linear kernel and optimized hyperparameters using an independent non-ALS dataset. The output, the nQiALS-Detection score, is a number between 0 and 1. Although the numeric score does not directly translate into a percent probability, higher scores can be interpreted as a higher likelihood that a given input from a typing sample belonged to a PALS. (Fig. 6).

Figure 6
figure 6

Schematic Overview of the nQiALS-Detection model. This model assesses ALS likelihood through typing pattern analysis. Input data from a 5-min semi-controlled typing stream generate statistics on normalized flight times. An SVM Classifier with a linear kernel, refined via sixfold cross-validation, produces a score between 0 and 1, indicating the likelihood of ALS.

Full size image

We defined experimental groups as follows: HC—those without ALS or other major neurological disease; ALS—all participants with ALS (PALS); mild ALS—PALS who scored ≥ 9 points in the ALSFRS-R fine motor subdomain (questions 4–7) at V1. We then performed two group comparisons: HC (n = 22) versus ALS (n = 49); HC (n = 22) versus Mild ALS (n = 17). Those with missing baseline typing data, insufficient information for nQiALS-Detection computation, or missing alternating finger tapping (AFT) results were excluded from the relevant analyses, resulting in smaller experimental subsets compared to the total participant pool.

The model’s performance was tested based on its ability to separate ALS samples from HC, quantified using the receiver operating characteristic area under the curve (ROC AUC). As a surrogate for early detection, performance was also evaluated using the ROC AUC built on the separation between HC and mild ALS. Additionally, we assessed Spearman ρ correlation between the model output and ALSFRS-R to evaluate the model’s ability to capture ALS-related functional impairment.

nQiALS-Progression

The nQiALS-Progression model was developed to estimate overall ALS progression rate, as measured by changes in total ALSFRS-R score, using the longitudinal stream of natural typing data. The model input consisted of a sequence of tensor-based representations of the weekly distributions of hold times, flight times, and Error X–Y. During training, the model utilized a numeric label as ground truth, representing the overall disease progression rate as measured by the total ALSFRS-R slope estimated over four clinical assessments (δALSFRS-R). The model was trained using a threefold stratified cross-validation approach to minimize overfitting the training folds. This experiment was conducted using only data from PALS. The resultant model is a recurrent neural network (RNN) that combined long short-term memory layers (LSTM) and fully connected deep layers, using a linear activation function to generate the network output. The output, the nQiALS-Progression score, is a number generally ≤ 0, as only declining ALS trends were represented in the training set. It can be interpreted broadly as the predicted slope of decline on the ALSFRS-R scale over the time interval represented by the typing input (Fig. 7). Input sequences of at least five weeks were required to generate a score, and each sequence was required to have at least 40% active weeks, and a consistency score exceeding 60% during the nine-month follow-up period. This consistency condition excluded sequences that presented excessive inactive weeks, employing a rolling window approach.

Figure 7
figure 7

Schematic Overview of the nQiALS-Progression model. This model estimates ALS progression rate using longitudinal typing data. Input comprises weekly distributions of hold times, flight times, and Error X–Y. A recurrent neural network (RNN) that integrates long short-term memory layers (LSTM) and fully connected layers, trained via threefold stratified cross-validation, predicts the ALSFRS-R slope (δALSFRS-R). The output score reflects the anticipated rate of ALS progression over the represented typing input interval.

Full size image

The model’s performance was evaluated using several statistical methods. Spearman correlation coefficients were first calculated to assess the rank correlation between the ALSFRS-R slope measured over the full follow-up period and the model's weekly scores. Receiver operating characteristic (ROC) analysis was performed to determine the area under the curve (AUC) scores at varying δALSFRS-R thresholds to separate faster vs slower progressors within the ALS group. Lastly, a Bland–Altman plot was used to visually compare the model's predicted average weekly scores with the reference δALSFRS-R, offering a view of the agreement between the predictions and the actual data.

nQiALS-Fine Motor

The nQiALS-Fine Motor model was developed to compare longitudinal typing patterns in participants with declining fine motor impairment over time to those with stable fine motor function as defined by the score on the fine motor subdomain of the ALSFRS-R. The model input consisted of a sequence of tensor-based representations of the weekly distributions of hold times, flight times, Error X/Y, velocity, and acceleration. During training, the model employed a binary label as ground truth to determine if a given sample represented a changing pattern (δFine motor ALSFRS-R < − 0.025/week, change) or a stable pattern (δFine Motor ALSFRS-R ≥ − 0.025/week, no change). The − 0.025 reference threshold was derived from the level of variability observed in healthy controls (HC), serving as an estimate of the inherent noise within the scale, given that HC were anticipated to maintain fine motor stability over time. A threefold stratified cross-validation approach was used to train the model, minimizing overfitting of the training folds. The final architecture of the model is a RNN that combines LSTM and fully connected deep layers, utilizing a sigmoid activation function to generate the network output. Input sequences of at least five weeks were required to generate a score, and each sequence was required to have at least 40% active weeks, and a consistency score exceeding 60% during the nine-month follow-up period.

The model employed a classification approach, due to the less granular nature of the fine motor scores on the ALSFRS-R scale and the relatively limited observable changes in this subdomain. Unlike the detection model, this classifier was recalibrated on a weekly basis, as it continually refined its predictions with the incorporation of new data. The model’s output is a number between 0 and 1, which can be interpreted as an indicator of the relative likelihood of a participant's ALSFRS-R fine motor subdomain score declining within the observed interval. By including the HC data, which is comparable to the ALS data with no change in the Fine Motor group, we were able to increase the sample size for this analysis (Fig. 8).

Figure 8
figure 8

Graphical representation of nQiALS-Fine Motor model. This model analyzes longitudinal typing patterns to differentiate between stable and declining fine motor function. Input includes weekly distributions of hold times, flight times, Error X/Y, velocity, and acceleration. A recurrent neural network (RNN) that integrates long short-term memory layers (LSTM) and fully connected layers, refined using threefold stratified cross-validation, provides a score between 0 and 1 indicating the likelihood of fine motor score decline within the observed interval.

Full size image

δFine Motor ALSFRS-R is defined as the weekly slope measured in the aggregate score of fine motor items of the ALSFRS-R (Q4–Q6), calculated using linear regression over the four visits. This metric provides an estimate of the progression of fine motor impairment in PALS. The threshold for differentiating between changing and stable patterns was established based on the minimum change measured in the HC group. This threshold is a decline in 0.025 points per week (which would correspond to < 1-point decline over the study duration of 9 months). By setting this threshold according to the degree of noise observed in HC, the model aimed to disregard non-ALS-related changes, allowing for a more accurate assessment of decline in fine motor performance due to ALS.

The experimental groups were organized as follows: HC (n = 23), ALS no change [δFine motor ALSFRS-R ≥ − 0.025 points/week (n = 26)], and ALS change [δFine Motor ALSFRS-R < − 0.025 points/week (n = 22)].

The model’s performance in early detection of fine motor decline was evaluated based on its ability to separate declining and stable typing trends, quantified using the ROC AUC.