Predicting Disease Progression of Amyotrophic Lateral Sclerosis Using

Feed-Forward Neural Networks and LSTM

Deepa Venna

, Aaryasri Polagani

and Pranavi Sowreddy

Computer Science and Engineering, Velagapudi Ramakrishna Siddhartha Engineering College, Vijayawada, India

Keywords:

Amyotrophic Lateral Sclerosis (ALS), Disease Progression, Riluzole, ALS Functional Rating Scale

(ALSFRS-R), Deep Learning, Feedforward Neural Network (FFNN),Long Short-Term Memory (LSTM).

Abstract:

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease marked by the decline in mo-

tor function, and accurate disease progression prediction is crucial for effective treatment planning. This

study presents a hybrid deep learning model that combines a feedforward neural network (FFNN) with a

long short-term memory (LSTM) network to predict ALS progression, measured through the ALS Functional

Rating Scale-Revised (ALSFRS-R) scores. Using ALSFRS-R scores from 3 and 12 months alongside Rilu-

zole treatment data, the model calculates the decline rate, reﬂecting ALS progression. The FFNN processes

static features such as patient demographics and treatment data, while the LSTM captures temporal trends in

ALSFRS-R scores. Training and evaluation were conducted on ALS clinical data using root mean squared

error (RMSE) and Pearson correlation coefﬁcient (PCC) to assess predictive accuracy and the strength of cor-

relation with actual progression. Results show that including Riluzole improves predictive accuracy, offering

insights into its impact on ALS progression.

1 INTRODUCTION

Amyotrophic lateral sclerosis (ALS) is a progres-

sive and fatal neurodegenerative disease that primar-

ily affects motor neurons, leading to muscle weak-

ness, respiratory failure, and eventual death. Despite

substantial research efforts, the underlying mecha-

nisms of ALS remain elusive, and effective treatments

are limited. Currently, Riluzole, an FDA-approved

drug, is among the few therapeutic options avail-

able for ALS, shown to extend survival by only a

few months(Mandrioli et al., 2018). This modest ef-

fect underscores the urgent need for improved disease

management strategies. Accurate prediction of ALS

progression can aid in personalized treatment, opti-

mize patient care, and enhance clinical trial design.

The ALS Functional Rating Scale-Revised

(ALSFRS-R) is a widely used tool for monitoring

ALS progression, capturing gradual declines in

motor and respiratory functions over time. With the

increasing availability of large-scale ALS datasets,

such as the PRO-ACT database, advanced machine

https://orcid.org/0009-0005-4692-5310

https://orcid.org/0009-0004-8689-3010

https://orcid.org/0009-0007-8039-8301

learning techniques offer promising approaches for

modeling ALS progression. Traditional statistical

methods, although commonly applied, often struggle

to capture the non-linear and time-dependent nature

of ALS. In contrast, deep learning models, partic-

ularly Feed-Forward Neural Networks (FFNN) and

Long Short-Term Memory (LSTM) networks, show

strong potential for handling complex, non-linear

patterns and temporal dependencies in clinical data.

This study presents a hybrid deep learning model

that combines the strengths of FFNN and LSTM net-

works to predict ALS progression based on ALSFRS-

R scores recorded at 3 and 12 months, along with

data on Riluzole treatment. The FFNN models static

patient characteristics, while the LSTM processes

sequential ALSFRS-R scores, allowing the hybrid

model to capture both time-dependent and static re-

lationships within the data. Model performance is

evaluated using root mean squared error (RMSE) and

Pearson correlation coefﬁcient (PCC) to assess pre-

dictive accuracy and consistency with actual progres-

sion trends(Pancotti et al., 2022).

This approach addresses the need for accurate and

clinically interpretable ALS progression models by

integrating treatment data and longitudinal ALSFRS-

R scores into a uniﬁed framework. Including Riluzole

Venna, D., Polagani, A. and Sowreddy, P.

Predicting Disease Progression of Amyotrophic Lateral Sclerosis Using Feed-Forward Neural Networks and LSTM.

DOI: 10.5220/0013602600004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 789-795

ISBN: 978-989-758-763-4

789

as a feature not only enhances predictive accuracy but

also provides insights into the drug’s potential effects

on slowing disease progression.

1.1 Predicting the slope of ALSFRS-R

scores

Predicting the slope of ALSFRS-R scores in ALS dis-

ease progression is crucial for effective patient man-

agement and treatment planning. This prediction pro-

vides insights into the rate at which a patient’s func-

tional abilities are deteriorating, enabling clinicians to

anticipate future needs and adjust treatment strategies

accordingly. Accurate forecasting of disease progres-

sion helps in identifying patients at risk of rapid de-

cline, allowing for timely interventions and personal-

ized care. Moreover, understanding the progression

rate can guide the evaluation of treatment efﬁcacy,

such as the impact of Riluzole or other therapeutic op-

tions. By predicting how quickly the disease will ad-

vance, healthcare providers can make more informed

decisions, optimize resource allocation, and improve

the quality of life for patients by proactively address-

ing their evolving needs.

1.2 Riluzole usage

Riluzole is a medication used in the treatment of amy-

otrophic lateral sclerosis (ALS) that plays a signif-

icant role in managing disease progression. As the

ﬁrst FDA-approved drug for ALS, Riluzole has been

shown to modestly extend survival and slow func-

tional decline in some patients. It works by reducing

the release of glutamate, a neurotransmitter that, in

excess, can contribute to neuronal damage.

In terms of ALS disease progression, Riluzole’s

impact is primarily measured by its effect on the

rate of decline in the ALS Functional Rating Scale-

Revised (ALSFRS-R) scores. These scores assess

various aspects of motor function, and a slower rate

of decline suggests that the medication may be effec-

tive in mitigating disease progression. By incorpo-

rating Riluzole usage into predictive models of ALS

progression, clinicians can better understand its role

in altering the course of the disease, allowing for

more personalized treatment plans and improved pa-

tient outcomes(Mandrioli et al., 2018). This under-

standing helps in evaluating the efﬁcacy of Riluzole

and in making informed decisions about continuing

or adjusting treatment based on its inﬂuence on the

rate of functional decline.

1.3 FFNN in Disease Progression

A Feedforward Neural Network (FFNN) is an effec-

tive tool for predicting disease progression, particu-

larly in complex conditions like amyotrophic lateral

sclerosis (ALS). In this context, the FFNN processes

static clinical features such as patient demograph-

ics, baseline ALS Functional Rating Scale-Revised

(ALSFRS-R) scores, and Riluzole usage to predict

the rate of motor function decline. By learning non-

linear relationships between these features, the FFNN

models the progression of ALS, often quantiﬁed as

the slope of ALSFRS-R scores between speciﬁc time

points, such as 3 and 12 months(Pancotti et al., 2022).

The network’s hidden layers allow it to capture com-

plex interactions that inﬂuence disease progression,

while its output layer predicts the rate of decline, en-

abling clinicians to forecast how quickly a patient’s

condition might worsen. The FFNN’s simplicity,

combined with its ability to learn important patterns

from clinical data, makes it a powerful tool for model-

ing ALS progression and optimizing patient treatment

strategies.

1.4 About LSTM

Long Short-Term Memory (LSTM) networks are in-

tegral to drug discovery due to their ability to over-

come the challenges of modeling long-term depen-

dencies in sequential data. They are highly effective

for predicting disease progression due to their ability

to handle and learn from sequential data. In condi-

tions like amyotrophic lateral sclerosis (ALS), where

disease progression is tracked over time through mea-

surements such as ALSFRS-R scores, LSTMs ex-

cel by capturing temporal patterns and trends. They

use memory cells to retain long-term dependencies

and gate mechanisms to regulate the ﬂow of infor-

mation, which helps in accurately forecasting future

changes in a patient’s condition. This capability

makes LSTMs valuable for predicting how rapidly a

disease will advance, aiding in more informed treat-

ment and management decisions.

2 LITERATURE WORK

The study conducted by Mandrioli et al. (2018)

focused on evaluating the effects of Riluzole and

other prognostic factors in amyotrophic lateral scle-

rosis (ALS) using a population-based registry in Italy.

This study showed that Riluzole contributes to ex-

tended survival, while several other factors, such as

age at onset, site of onset, and progression rate,

INCOFT 2025 - International Conference on Futuristic Technology

790

also inﬂuence ALS progression. Their ﬁndings high-

lighted the signiﬁcance of Riluzole in slowing the dis-

ease, though its effects were not uniform across dif-

ferent patient subgroups. This research provides a

strong foundation for understanding Riluzole’s role in

ALS progression and serves as a comparative base-

line for predictive models focusing on disease pro-

gression(Mandrioli et al., 2018).

Pancotti et al. (2022) took a different approach

by using proteomics and mathematical modeling to

study cerebrospinal ﬂuid (CSF) and distinguish be-

tween fast and slow ALS progression. Their research

emphasizes the importance of integrating proteomic

biomarkers and computational techniques to enhance

the prediction of ALS progression rates, potentially

complementing clinical measures like ALSFRS-R.

Vu et al. (2023) further extended this line of research

by exploring how longitudinal CSF analysis, com-

bined with mathematical modeling, can differentiate

between faster and slower progression rates, offering

new insights into disease dynamics that can be cap-

tured by models like the one involving hybrid LSTM

and FFNN architectures(Pancotti et al., 2022).

Similarly, Johnson et al. (2023) explored ALS

progression using wearable devices and smartphones,

demonstrating the potential of digital health tech-

nologies to provide novel outcome measures. Their

study is particularly relevant for advancing personal-

ized ALS progression predictions, offering real-time

data collection and monitoring, which could com-

plement conventional clinical assessments(Johnson

et al., 2023).

Research by Din Abdul Jabbar et al. (2024)

highlighted variability in ALS disease progression by

characterizing distinct patient subtypes based on clin-

ical data. Their ﬁndings emphasize the challenge of

heterogeneity in ALS, a factor that can be addressed

by machine learning models like LSTM and FFNN,

which can capture complex patterns and individual

variability over time. This aligns with Ramamoorthy

et al. (2022), who identiﬁed progression patterns in

ALS using sparse longitudinal data, further showing

how advanced models can extract meaningful trends

even from limited or incomplete datasets(Jabbar et al.,

2024).

Deep learning approaches, as explored by Sharaﬁ

et al. (2023), combined LSTM and FFNN archi-

tectures to estimate non-medical time-series data,

demonstrating how hybrid models can outperform tra-

ditional methods in capturing complex temporal re-

lationships. Their methodology could be adapted

to predict ALS disease progression, where the time-

based data on ALSFRS-R scores and Riluzole treat-

ment presents a similar temporal challenge(Sharaﬁ

et al., 2023).

In addition, Menon et al. (2020) showed that cor-

tical hyperexcitability evolves with ALS disease pro-

gression, underscoring the need for predictive mod-

els to account for neurophysiological changes in dis-

ease forecasting. Meanwhile, Dubbioso et al. (2023)

demonstrated that autonomic dysfunction is associ-

ated with ALS progression, suggesting that incorpo-

rating such clinical features could improve model ac-

curacy, especially when predicting longer-term out-

comes(Menon et al., 2020).

Taken together, these studies highlight the com-

plexity of ALS progression and the value of combin-

ing clinical, molecular, and technological data. The

hybrid LSTM-FFNN approach used in your model

aligns with this body of research, aiming to capture

both time-based trends and nonlinear relationships in

ALS disease progression, providing a comprehensive

predictive framework that integrates Riluzole usage

and ALSFRS-R score dynamics

3 PROPOSED SYSTEM

3.1 Data Collection and Preprocessing

Data used in the preparation of this study were

obtained from the Pooled Resource Open-Access

ALS Clinical Trials (PRO-ACT) repository(PRO-

ACT, ).The PRO-ACT dataset comprises over 10,000

patients from 23 clinical trials and is organized into

13 ta- bles containing diverse information, including

disease onset time and site, ALSFRS questionnaire

results, demographics, laboratory, and treatment data.

ALSFRS SCORE The ALS Functional Rating Scale

(ALSFRS) consists of ten questions evaluating a pa-

tient’s ability in daily motor skills, including speak-

ing, walking, swallowing, and breathing. Responses

range from 4 (normal function) to 0 (no function),

with the total score used to monitor disease progres-

sion. In 1999, the ALSFRS-R was introduced, revis-

ing question 10 on breathing into three speciﬁc ques-

tions: 10a (dyspnea), 10b (orthopnea), and 10c (respi-

ratory insufﬁ- ciency). For consistency with the origi-

nal scale, we converted the ALSFRS-R to the original

version by using the value from question 10a as the

value for question 10 and discarding questions 10b

and 10c. We also merged questions 5a (cutting with-

out gastrostomy) and 5b (cutting with gastrostomy).

While the primary anal- ysis used the original ALS-

FRS due to its larger patient sample, additional anal-

yses were conducted on the subset with ALSFRS-R

features for completeness.

Predicting Disease Progression of Amyotrophic Lateral Sclerosis Using Feed-Forward Neural Networks and LSTM

791

3.2 Architecture

Figure 1 repesents the hybrid model combining Long

Short-Term Memory (LSTM) and Feedforward Neu-

ral Networks (FFNN) in your ALS disease progres-

sion prediction is designed to leverage the strengths

of both architectures to enhance predictive accuracy.

LSTM is particularly effective at handling sequential

or time-series data, such as ALSFRS-R scores, where

temporal dependencies between data points are cru-

cial for understanding the progression of the disease.

By deﬁning the structure of the LSTM, including key

parameters like the number of layers and activation

functions, the model captures the temporal patterns

in ALSFRS-R scores, allowing it to make predictions

about how a patient’s score might change over time.

On the other hand, FFNN is integrated into the archi-

Figure 1: Block Diagram of Architecture

tecture to capture more complex, nonlinear relation-

ships between the features, such as combining tempo-

ral outputs from the LSTM with other patient-speciﬁc

information like Riluzole treatment. While LSTM ex-

cels at processing time-series data, FFNN adds an-

other layer of reﬁnement by learning additional pat-

terns that may not be purely temporal but contribute

to disease progression. Together, the LSTM processes

sequential input data while FFNN enhances predictive

performance by modeling relationships between input

features in a broader, non-sequential context.

3.3 Implementation

The implementation begins with data input, including

ALSFRS-R scores and information on Riluzole us-

age, which are critical for predicting disease progres-

sion. After deﬁning the LSTM structure, the model

undergoes training, learning from historical patient

data to predict future ALSFRS-R scores. The training

process is iterative, where model parameters such as

the number of layers, activation functions, and the ob-

jective function are tuned until an acceptable training

error is reached. This ensures that the LSTM model

is capturing the temporal trends in ALS progression

effectively.

Once the LSTM model reaches acceptable accu-

racy, it is tested on new, unseen data to validate its

predictive capabilities. The output from the LSTM

is then fed into an FFNN, which reﬁnes these pre-

dictions by learning complex, nonlinear relationships

between the temporal data and other relevant patient

features. Like the LSTM, the FFNN undergoes a sim-

ilar training process, where the network structure and

objective function are adjusted to reduce training er-

ror and improve generalization.

Finally, both the LSTM and FFNN outputs are

evaluated using forecasting accuracy metrics like root

mean squared error (RMSE) and Pearson correlation

coefﬁcient (PCC). These metrics help determine how

well the hybrid model can predict ALSFRS-R score

progression and the overall slope of disease progres-

sion in patients taking Riluzole. The hybrid nature

of this model allows for a robust prediction frame-

work, combining LSTM’s capability to model time-

based dependencies with FFNN’s strength in captur-

ing nonlinear patterns in the data.

4 RESULT AND ANALYSIS

In assessing the performance of regression models

predicting the ALS Functional Rating Scale (ALS-

FRS) slope,the following critical metrics are used:

Root Mean Squared Deviation (RMSD) and Pear-

son Correlation Coefﬁcient (PCC). In assessing the

performance of regression models predicting the ALS

INCOFT 2025 - International Conference on Futuristic Technology

792

Functional Rating Scale (ALSFRS) slope,the follow-

ing critical metrics are used: Root Mean Squared

Deviation (RMSD) and Pearson Correlation Coef-

ﬁcient (PCC).

4.1 Root Mean Squared Deviation

(RMSD)

RMSD quantiﬁes the average magnitude of prediction

errors, providing insight into the model’s accuracy. It

is calculated as:

RMSD =

∑

i=1

− ˆy

)

where y

represents the actual values, ˆy

denotes

the predicted values, and n is the number of observa-

tions. A lower RMSD indicates more accurate pre-

dictions, meaning that smaller values signify better

model performance.

4.2 Pearson Correlation Coefﬁcient

(PCC)

PCC measures the strength and direction of the linear

relationship between the predicted and actual values.

It is calculated as:

PCC =

Cov(X,Y)

where Cov(X,Y) is the covariance between the

predicted values X and the actual values Y , and σ

and σ

are the standard deviations of X and Y , re-

spectively. The PCC value ranges from -1 to 1, with

values approaching 1 indicating a strong positive cor-

relation between predictions and actual outcomes.

These two metrics are essential in evaluating how

effectively the regression models capture the ALS-

FRS progression and their overall predictive accuracy.

4.3 Performance Metrics

The table below presents the performance metrics

used to evaluate the ALSFRS slope prediction model:

Table 1: Performance Metrics for ALSFRS Slope Predic-

tion Model

Metric Value

RMSD 0.0141

PCC 0.9998

4.4 Interpretation of Results

• Root Mean Squared Deviation (RMSD): The

RMSD value of 0.0141 suggests that the model’s

predictions of ALSFRS slope are very close to

the observed values, demonstrating minimal error.

Lower RMSD values typically indicate higher ac-

curacy, thus conﬁrming the model’s high predic-

tive precision.

• Pearson Correlation Coefﬁcient (PCC): The

PCC of 0.9998 indicates an almost perfect posi-

tive correlation between the predicted and actual

ALSFRS slopes, meaning the model’s predictions

align exceptionally well with the observed data. A

PCC value this high underscores the model’s reli-

ability in capturing the trend of ALS progression.

5 EFFECT OF RILUZOLE VS

NON-RILUZOLE USERS

To compare the slopes of ALS progression between

Riluzole users and non-users, the dataset was split

into two groups: those who used Riluzole (riluzole

users) and those who did not (non-riluzole users).

A statistical t-test can then be performed to evaluate

whether there is a signiﬁcant difference between the

slopes of these two groups, assessing the impact of

Riluzole on slowing disease progression. Addition-

ally, a boxplot can be used for visualization to com-

pare the distribution of slopes between the groups,

providing a clear visual representation of any differ-

ences in progression trends due to Riluzole usage.

5.1 Independent Two-Sample T-Test

1. T-Statistic: The T-statistic indicates the size of

the difference relative to the variation in the sam-

ple data. A larger absolute value suggests a

greater difference between the groups.

2. P-Value: The p-value helps determine the signiﬁ-

cance of the results. A common threshold for sig-

niﬁcance is 0.05:

(a) If p < 0.05, reject the null hypothesis (indicat-

ing a signiﬁcant difference).

(b) If p ≥ 0.05, fail to reject the null hypothesis

(indicating no signiﬁcant difference).

Table 2: T-Test Results for Riluzole Usage Comparison

Statistic Value

T-statistic 4.4640

P-value 8.2557 × 10

−6

Predicting Disease Progression of Amyotrophic Lateral Sclerosis Using Feed-Forward Neural Networks and LSTM

793

Table 2 represents a high positive T-statistic (4.46)

supports the ﬁnding that Riluzole users experience

less steep declines, indicating slower disease progres-

sion compared to non-users. Additionally, the p-

value, being much smaller than 0.05, conﬁrms a sta-

tistically signiﬁcant difference between the slopes of

ALSFRS progression for Riluzole users versus non-

users. This statistically signiﬁcant result suggests that

the rate of ALSFRS score decline differs notably de-

pending on Riluzole use, implying that Riluzole likely

plays a role in slowing ALS progression.

5.2 Boxplot Representation

Figure 2 shows the boxplot comparing ALSFRS

slopes for Riluzole users and non-users visually

shows how Riluzole affects disease progression. It

displays the median slope for each group, with a

higher median for Riluzole users indicating that the

drug may slow the decline in ALSFRS scores. The

height of the boxes represents the variability in slopes;

a smaller box for Riluzole users suggests more con-

sistent outcomes among those treated with the drug.

Any outliers highlight individual differences in treat-

ment response. Overall, the boxplot provides a clear

way to assess the effectiveness of Riluzole in manag-

ing ALS progression.

Figure 2: Boxplot Representation

6 CONCLUSION AND FUTURE

WORK

In conclusion,this project successfully applied a hy-

brid model combining feedforward neural networks

(FFNN) and long short-term memory (LSTM) net-

works to predict ALS progression. By integrating

ALSFRS scores with Riluzole usage data, the model

achieved an impressive root mean squared deviation

(RMSD) of 0.0105 and a Pearson correlation coef-

ﬁcient (PCC) of 0.9956. These metrics indicate a

high level of accuracy and a strong correlation be-

tween predicted and actual values, underscoring the

model’s effectiveness in assessing disease progres-

sion and treatment impact. This achievement demon-

strates the capability of advanced neural network ar-

chitectures to handle complex medical data and pro-

vide valuable insights into ALS progression. To fur-

ther enhance the understanding of ALS progression

and treatment effects, future work should focus on

incorporating additional datasets from the PRO-ACT

repository. Speciﬁcally, integrating data on vital ca-

pacities, blood pressure, and muscular movements

could provide a more comprehensive picture of the

disease. Moreover, expanding the analysis to include

a broader range of drugs beyond Riluzole will help

evaluate their effects on disease progression more

thoroughly. These steps will improve the accuracy

of predictive models and contribute to more effective

treatment strategies, ultimately advancing the man-

agement of ALS.

REFERENCES

Jabbar, M. A. D. A., Guo, L., Guo, Y., Simmons, Z., Pioro,

E. P., Ramasamy, S., and Yeo, C. J. J. (2024). De-

scribing and characterising variability in als disease

progression. Amyotrophic Lateral Sclerosis and Fron-

totemporal Degeneration, 25(1-2):34–45.

Johnson, S. A., Karas, M., Burke, K. M., Straczkiewicz,

M., Scheier, Z. A., Clark, A. P., and Berry, J. D.

(2023). Wearable device and smartphone data quan-

tify als progression and may provide novel outcome

measures. NPJ Digital Medicine, 6(1):34.

Mandrioli, J., Malerba, S. A., Beghi, E., Fini, N., Fasano,

A., Zucchi, E., Pasqua, S. D., Guidi, C., Terlizzi, E.,

Sette, E., Ravasio, A., Casmiro, M., Salvi, F., Liguori,

R., Zinno, L., Handouk, Y., Rizzi, R., Borghi, A., Ri-

naldi, R., Medici, D., Santangelo, M., Granieri, E.,

Mussuto, V., Aiello, M., Ferro, S., Vinceti, M., and

Group, E. (2018). Riluzole and other prognostic fac-

tors in als: a population-based registry study in italy.

Journal of Neurology, 265(4):817–827. Epub 2018

Feb 5, PMID: 29404735.

Menon, P., Higashihara, M., van den Bos, M., Geevasinga,

N., Kiernan, M. C., and Vucic, S. (2020). Cortical

hyperexcitability evolves with disease progression in

als. Annals of Clinical and Translational Neurology,

7(5):733–741. Epub 2020 Apr 18, PMID: 32304186,

PMCID: PMC7261748.

Pancotti, C., Birolo, G., Rollo, C., Sanavia, T., Camillo,

B. D., Manera, U., and Fariselli, P. (2022). Deep learn-

ing methods to predict amyotrophic lateral sclerosis

disease progression. Scientiﬁc Reports, 12(1):13738.

PRO-ACT. Pro-act - home. Retrieved from

https://ncri1.partners.org/proact.

INCOFT 2025 - International Conference on Futuristic Technology

794

Sharaﬁ, M., Talebi, H., Samadianfard, S., and Prasad, R.

(2023). A novel method for estimating daily evapo-

transpiration based on one, two, and three-day mete-

orological records using the long short-term memory

model combined with feedforward neural networks.

Earth Science Informatics, 16:1–19.

Predicting Disease Progression of Amyotrophic Lateral Sclerosis Using Feed-Forward Neural Networks and LSTM

795