Indonesian J our nal of Electrical Engineering and Computer Science V ol. 42, No. 3, June 2026, pp. 835 ∼ 845 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v42.i3.pp835-845 ❒ 835 Sensor -based pr ediction of ALS pr ogr ession: exploring PHI and featur e engineering Chib uzor Chukwuemeka Ok er e, Ed win Thuma, Gontlafetse Mosweunyane Department of Computer Science, Uni v ersity of Botsw ana, Gaborone, Botsw ana Article Inf o Article history: Recei v ed Jul 26, 2025 Re vised Mar 25, 2026 Accepted May 26, 2026 K eyw ords: Amyotrophic lateral sclerosis Feature engineering Machine learning Ph ysical health inde x Sensor data ABSTRA CT Amyotrophic lateral sclerosis (ALS) is a serious disease that af fects nerv e and muscle function, with no kno wn cure. Early and accurate monitoring is essen- tial to help ph ysicians pro vide better care. Although machine learning has been applied to pre dict the progression of ALS, man y models struggle with issues such a s poor data quality and missing information, which af fect accurac y . In this paper , our aim is to impro v e e xisting models by introducing bet ter features to enhance prediction performance. A k e y contrib ution is the de v elopment of a ne w feature called the ph ysical health inde x (PHI), which combines four im- portant patient attrib utes: body mass inde x (BMI), weight, forced vital capacity (FVC), and basal calories. This feature pro vides a clearer vie w of the ph ysical health of the patie nt, enabling the model to learn more ef fecti v ely . W e used the IDPP CLEF 2024 BT O dataset and performed three e xperiments: using 50 ra w features, 29 enginee red features, and 25 further engineered features including PHI. The results sho wed that the R-squared of the XGBoost model impro v ed from 0.9573 to 0.9663 and nally 0.9828, while RMSE decreased from 0.2317 to 0.1801 and then 0.1182 with PHI. This study hi ghlights ho w tar geted feature engineering can impro v e the prediction of ALS using machine learning. This is an open access article under the CC BY -SA license . Corresponding A uthor: Chib uzor Chukwuemeka Ok ere Department of Computer Science, Uni v ersity of Botsw ana Notw ane Road, 4775, Gaborone, Botsw ana Email: moeok ere17@yahoo.com 1. INTR ODUCTION Amyotrophic lateral sclerosis (ALS) is a progressi v e neurode generati v e disease of adulthood caused by the loss of spinal, b ulbar , and cortical motor neurons, which can cause v oluntary muscles within the human body to paralyze [1]. The e xact cause of this disease remains unkno wn, and a permanent cure for this disease is yet to be disco v ered, although there are some drugs a v ailable that can be used to slo w the progression of the disease in patients. The dif culty in identifying the main cause of ALS is attrib uted to the v ariation in the progression of the disease [2]. Due to this dif culty , an approach has been de v eloped to track and e xamine the progression of the disease in patients [3]. The strength testing e v aluation w as de v eloped to quantify the strength of an ALS patient, as weakness is a major clinical f actor of the ALS disease [4]. F orced vital capacity (FVC) w as de v eloped to measure the full breathing capacity of a patient by considering the respiratory functioning of a patient [5]. The ALS functional rating scale (ALSFRS) is the tool commonly used to measure the progression of ALS in patients [1], [4]. It is a scale-based questionnaire that consists of 10 questions that are usually gi v en to ALS patients to e v aluate specic functional loss in acti vities such as w alki ng, climbing, speaking, sw allo wing, and autonomy in dressing [1]. J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
836 ❒ ISSN: 2502-4752 In 1999, the ALSFRS e v aluation questions were re vised and are no w called ALSFRS-R and no w contain 12 e v aluat ion questions that also tak e the respiratory functions of patients into consideration [5], [6]. The ALSFRS-R has a scale ranging from 0 to 4 for each question, where 0 is the w orst and 4 is t he normal scale. W ith these scale v alues, the highest total ALSFR S-R score a patient can obtain is 48, signifying a health y condition, while the lo west score is 0, signifying the w orst condition [7]. Ov er the years, obtaining these scores manually has become error prone due to human mistak es making the ef fecti v eness of the approach questionable. Adv ances in technology ha v e brought se v eral adv anced t ools and techniques, such as machine learning and st atistical approaches, to the eld of ALS prediction [2]. Machine l earning algorithms ha v e enabled researchers to predict the slope of ALS, which can be used to e v aluate the progression of the disease [1], [7] to identify ALS patients’ risk groups [4], [8], and to predict the se v erity of ALS in patients [7]. Earlier studies ha v e used tw o dif ferent machine learning approaches, supervised and unsupervised learning, to predict the se v erity of ALS in patients. Supervised learning approaches use techniques such as classication, re gression, and deep learning on patient-labelled data to perform the prediction of ALS pro- gression in patients [1] while unsupervised learning uses approaches such as association and clustering [6]. Using machine learning approaches has impro v ed the accurac y , e v aluation, and ef fecti v eness of the ALSFRS- R scales, as it has helped reduce the errors associa ted with human in v olv ement i n data collection [9]. Despite the ef cienc y of machine learning algorithms in the eld o f ALS prediction, the high le v el of heterogeneity among patients has made the de v elopment of generalized and reliable models dif cult [2], [10]. Hence, the need to de v elop reliabl e machine learning models is important to the eld of ALS prediction, as it w ould help in tracking the progression of the disease more ef ciently . Se v eral researchers ha v e used the pooled resource open-access clinical trials (PR O-A CT) ALS datas et to de v elop machine learning models for ALS progression prediction [1], [4], [7]. T urabieh et al. [7] deplo yed a con v olutional deep neural netw ork (CDNN) to de v elop a model that can predict the se v erity of ALS disease. Their model produced an A UC score of 86%. Ho we v er , this study focuses on de v eloping a machine-learning model that can e xamine only the v oice recordings (speech functions) of patients; hence, the ndings of the study are limited. P ancotti et al. [1] used se v eral deep neural netw ork (DNN) algorithms to predict the slope of ALS prediction in patients using the PR O-A CT dataset. The authors pointed out that using DNN algo- rithms to predict ALS is still a ne w area and needs more research. T o perform the prediction of ALS, the authors de v eloped 6 predicti v e models using machine learning algorithms, of which four were neural netw orks con v olutional neural netw ork (CNN), feed-forw ard neural netw ork (FFNN), recurrent neural netw ork (RNN), FFNN+CNN) and the other tw o were traditional machine learning models bidirectional and auto-re gressi v e transformer (B AR T) and random forest (RF). The authors measured the performance of the de v eloped models using root mean squared de viation (RMSD) and pearson correlation coef cient (PC C), for which FFNN+CNN outperformed other models as it produced the least RMSD and PCC scores of 0.543 and 0.415, respecti v ely . The study , ho we v er , is limited as only a limited number of observ ations were e xamined in making ndings; hence, the performance of the model cannot be generalized. Gordon and Lerner [6] proposed a machine-learning algorithm which is capable of e xamining the 1- year survi v al rate and disease progression of ALS patients. Similar to [1], the authors utilised the PR O-A CT dataset where the y performed feature selection using the ridge re gression (RR) model. T o handle missing v alues, the authors dropped all columns containing missing v alues, hence reducing the size of the dataset. T o de v elop the models, both supervised and unsupervised learning approaches were applied. F or the supervised learning, a light gradient boosting machine (LGBM) w as used, while for the unsupervised learning, Uniform manifold approximation projection (UMAP) w as used. The de v eloped prediction model using Supervised learning had an RMSE of 2.86 and an R-squared v alue of 0.79, while the de v eloped prediction model using unsupervised learning had a prediction accurac y of 92.75%. The limited data size w as due to the dropping of missing v alues therefore limiting the ndings of this study . T asks similar to this w ork and util izing the same dataset were carried out as part of the the int elligent disease progression prediction (iDPP) Challenge at the 2024 Conference and Labs of the Ev aluation F orum (CLEF 2024) [11]. Marinello et al. [12] applied both linear (LR and RR) and non-linear (RF) models to predict ALSFRS-R scores in ALS patients using sensor and en vironmental data. The system w as trained using the backw ard feature selection technique, starting with all features and then remo ving them one by one, with the nal model trained with the susbset of features that resulted in the lo west root mean square error (RMSE) score. When using sensor data the approach sho wed impro v ement in the model’ s performance. The RMSE o v er the predicted 12 ALSFRS-R scores impro v ed from [0.463-0.733] to [0.286-0.582]. Martins et al. [13] applied Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 835–845 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 837 temporal summarization techniques to the highly dimensional ALS sensor data, then used feature selection or e xtraction techniques (K-Best selection across all questions, K-Best selection by question, and biclustering) to obtain a representati v e statistics with smaller dimensionality . The y then utilised the scikit-learn algorithms (LR, RF , XGBoost, and support v ector machines (SVM)) to predict the ASLFRS-R scores with each predicting an ASLFRS-R question. The best results were yielded by RF with a RMSE of 0.676 for one task and 0.530 for another , although it did not outperform the baseline results for most of the questions. The authors concluded that there w as no consensus re g arding the best feature selection or e xtraction approach, and that further research w as needed on capturing the temporal patterns of sensors to understand their potential in tracking the progress of ALS. Mehta et al. [14] also compressed and amalg amated the temporal ALS sensor data using statistical methods before applying machine learning models to predict the ASLFRS-R scores. Features were generated from the data using techniques such as median aggre g ation, feature-based time series analysis, and long short- term memory (LSTM) neural netw orks. V arious machine learning models including traditional ML models and deep learning models were then e xperimented with to predict the ALSFRS-R scores. The nai v e bayes classier and the Elastic Net, a re gularized linear re gres sion model produced the best results, with the nai v e bayes model achie ving a MAE of 0.20 and a RMSE of 0.49, slightly outperforming the Elastic Net model, which recorded an MAE of 0.22 and a RMSE of 0.50. The authors ho we v er posit that the Elastic Net+Nai v e re gression model combination of fered a rob ust frame w ork for understanding feature contrib utions. Silv a and Oli v eira [15] used the same ALS sensor dataset and used statistical measures such as mean, median, standard de viation, minimum, maximum, and range to aggre g ate the data. Feature selection w as then applied based on rele v ance and signicance and preliminary models were trained to e v aluate the importance of each feature, therefore helping to identify , select and use only f eatures which helped most with the predicti v e accurac y of the models. The Ensemble learning method, RF clas sier , w as then used for the prediction of ALSFRS-R due to its ability to handle missing labels in classication. The best model utilized temporal analysis and achie v ed a MAE of 0.25 in one task and 0.326 in another , with RMSE of 0.544 and 0.608 respecti v ely . These studies sho wed that machine learning algorithm s can be ef fecti v e in measuring the progress of ALS within its patients, enabling quick treatment and slo wing do wn the impact of the sickness on patients, as insights deri v ed from machine learning processes w ould help in making informed decisions related to the well-being of patients. Zhou and Manser [16] in their study ,indicated that predicting ho w ALS will progress should not just focus on patients in general. The y stressed that it is also v ery important to create models that can help identify patients who may not li v e for long. This w ould help both patients and doctors gi v e the right treatments to e xtend patients’ li v es. Because of this, the y used machine learning algorithms to b uild m odels that c an predict which patients may ha v e a shorter survi v al time. T o achie v e this, the y used an Ensemble imbalance method along with six common machine learning algorithms on the PR O-A CT dataset. These algorithms included K- nearest neighbour (KNN), decision tree (DT), RF , SVM, Nai v e Bayes, and NN. In their empirical e v aluation of the models the y de v eloped, the neural netw ork produces the best performance with an accurac y of 88%. Ev en though the models were quite accurate, the authors pointed out that more adv anced algorithms, often called ”black box es” lik e neural netw orks, should be considered. These adv anced methods could help create stronger and more reliable models compared to traditional ones. The study conducted by Ok ere , Thuma and Mosweun yane [17] disagrees with this as DNN models de v eloped in this study were outperformed by ensemble models, sho wcasing the limitations of DNNs. Ho w- e v er , the results achie v ed in this study may be inuenced by o v ertting due to absence of rigorous data pre- processing such as handling missing v alues and feature engineering. Hence, the authors highlighted the need for ef fecti v e data pre-processing suggesting that the use of adv anced techniques such feature engineering and proper handling of missing v alues can help in achie ving better performance and de v elopment of more reliable machine learning models [17]. This study seeks to address this g ap by introducing feature engineering and data bootstrapping on what w as pre viously done in [17]. Unlik e e xisting ALS predict ion studies that rely on ra w clinical v ariables or high- dimensional sensor features independently , this study introduces a composite ph ysical health inde x (PHI) that inte grates respiratory capacity (FVC), body mass inde x (BMI) and weight, and metabolic demand (calories) into a single interpretable feature. While pre vious w orks ha v e used these v ari ables in isolation, none ha v e combined them into a unied inde x designed to capture the o v erall ph ysiological resilience of ALS patients. This mak es PHI a no v el feature that bridges clinical rele v ance and machine learning usage. In addition, this study uses sensor data alongside se v eral machine learning techniques such as data Sensor -based pr ediction of ALS pr o gr ession: e xploring PHI and ... (Chib uzor Chukwuemeka Ok er e) Evaluation Warning : The document was created with Spire.PDF for Python.
838 ❒ ISSN: 2502-4752 transformation, data bootstrapping, feature engineering, feature selection and feature scaling to de v elop se v eral predicti v e machine learning models for monitoring ALS progression to impro v e what w as done in the pre vious studies [16], [17]. Feature Engineering w as used to create a ne w feature called PHI from the combination of 4 major patient data which were BMI, calories, weight, and FVC. The chosen features BMI, calories, weight, and FVC collecti v ely of fer a well-rounded vie w of a n indi vidual’ s ph ys ical health by addressing body composition, nutritional habits, and respiratory function, which is vital in analyzing the present ALS condition of a patient [18]. The combination of these features ensures that the PHI reects both metabolic and cardiopulmonary well-being for a more comprehensi v e health assessment for ALS patients [19]. The remaining part of the paper proceeds as follo ws: Section refmethiod outlines the method used for model de v elopment. Section refresults-d presents the results obtained and discussions on these results, while section 4. outlines the conclusions of this study based on ndings made and also points out areas for impro v ement in future w ork. 2. METHOD 2.1. Data description The data utilized in this study w as secondary data obtained from the IDPP CLEF 2024 competi- tion [11], which w as pro vided by the BRAINTEASER ontology (BT O) team. There were 6 distinct datasets: 2 static dat a les, 2 sensor data les, the clinic ALSFRS scores and the self assessed ALSFRS-R score le. These were mer ged into a single le to enable the creation of the ne w features used in this study . T o mer ge the datasets, the P andas library w as used to mer ge them into a single P andas data frame to aid model de v elopment. The patient ID w as used a longside the days of diagnosis to combine the 6 datasets. The ne wly formed dataset has 486 patient records and 110 attrib utes; ho we v er , the dataset has 14040 missing v alues, which is relati v ely high and needed to be handled properly (later in section 2.2. we use auto-encoders and multiple imputation to handle missing v alues). After pre-processing the dataset and performing feature engineering, the ne wly formed dataset had 49 features, of which 12 were the v alues to be predicted (Q1 – Q12), while the remaining 37 were patient data formed from the static and feature-engineered sensor dataset. T able 1 sho ws the data attrib utes and their descriptions. 2.2. Data exploration Data e xploration w as done in this study to help in the identication of lapses within the dataset, e.g. missing v alues, as the sensor data to be used contained a lot of missing v alues (14040) due to patients not pro viding their details daily [20]. The e xploration conducted sho wed that the minimum age of a patient in the dataset w as 21 years, while the maximum age w as 73. Ho we v er , e xploration of the dataset using the describe() function sho wed that the a v erage age of ALS patients in the dataset w as 56 years. Further e xploration using visualisation w as done to sho w the distrib ution of ALS sta tes within the dataset. T o do so, feature engineering w as done to group patients based on their age, and then the ALSFRS-R scores were used to group these patients into 4 distinct cate gories of ALS status: V ery bad, Bad, F air , and Good, based on the status scale pro vided by [20]. The tw o missing v alue techniques used were Auto-Encoders and Multiple Imputation. 2.3. F eatur e engineering In this study , feature engineering w as performed on the sensor dataset, which contains sensor data of patients holding records of their steps, respiration, heart rates, and heartbeats. Since there are man y features in the dataset, to enhance the performance of the model, se v eral descripti v e statistics such as mean, median, standard de viation, minimum, and maximum v alues were used to form ne w features on the dataset to pro vide unied data for better performance. T o perform feature engineering, all similar columns present within the sensor dataset were selected from the dataset and placed in a unique list. 5 distinct lists were formed, holding the names of columns that ha v e similar records. The lists formed were beat columns, heart rate columns, respi- ration columns, spo2 columns and steps columns. The selected columns were used to form ne w features in the dataset by unifying similar columns present in each formed list to get their mean, median, standard de viation (std de v), maximum v alues and minimum v alues for each patient. Furthermore, the PHI feature w as created by nding the a v erage of the weight, calories, FVC and BMI. This w as signicant to help pro vide a simplied b ut rob ust representation of the static and sensor data of patients, which can be utilised to optimise the perfor - mance of models de v eloped. Upon creating these ne w features, columns engineered to form ne w columns were remo v ed from the data set to a v oid redundanc y . Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 835–845 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 839 T able 1. Data features description No. Feature Description 1 Days T ime since diagnosis, e xpressed in days 2 acti v e calories T otal acti v e calories in a patient 3 basal calories T otal basal calories in a patient 4 total calories The sum of acti v e and basal calories 5 total steps T otal steps within the considered se gment of data 6 beat to beat mean A v erage RR interv al v alue of a patient’ s heartbeat 7 beat to beat median Median RR interv al v alue of a patient’ s heartbeat 8 beat to beat std The standard de viation of the RR interv al v alues 9 beat to beat min Minimum RR interv al v alue of a patient’ s heartbeat 10 beat to beat max Maximum RR interv al v alue of a patient’ s heartbeat 11 heart mean The mean heart rate v alue of a patient 12 heart median Median heart rate v alue of a patient 13 heart std The standard de viation of heart rate v alues 14 heart min The minimum heart rate v alue of a patient 15 heart max The maximum heart rate v alue of a patient 16 res piration mean The mean respiratory rate of a patient 17 res piration median Median respiratory rate of a patient 18 respiration std The standard de viation of respiratory rates 19 res piration min Minimum respiratory rate of a patient 20 res piration max The maximum respiratory rate of a patient 21 spo2 mean Mean oxygen saturation le v el (SpO2) of a patient 22 spo2 median Median pulse oximetry of a patient 23 spo2 std Standard de viation of SpO2 v alues 24 spo2 min The minimum pulse oximetry v alue of a patient 25 spo2 max The maximum pulse oximetry v alue of a patient 26 steps mean A v erage number of steps w alk ed by a patient 27 steps median Median of steps w alk ed by a patient 28 steps std Standard de viation of steps w alk ed 29 steps min Minimum number of steps w alk ed 30 steps max Maximum steps w alk ed by a patient 31 Se x Gender of the patient 32 di agnostic delay Delay between ALS onset and diagnosis 33 age at diagnosis P atient’ s age at diagnosis 34 FVC F orced V ital Capacity at diagnosis 35 W eight W eight measured at diagnosis 36 Bmi Body Mass Inde x at diagnosis 37 Id P atient ID (anon ymized; dropped for pri v ac y) 38 Q1 – Q12 ALSFRS-R scores from patient questionnaires 2.4. Data transf ormation Other pre-processing steps done include data encoding and feature selection. F or data encoding, the label encoder pro vided by the Sklearn preprocessing library [21] w as utilised to con v ert cate gorical data present within the dataset (typically se x) to numerical v alues to ensure all data are in the appropriat e format, which can be easily unders tood by machine learning algorithms during model de v elopment. Ne xt, feature scaling through normalisation and standardisation of the cleaned dataset w as carried out using the MinMaxScaler and the StandardScaler libraries pro vided by the Sklearn preprocessing library . 2.5. Data sampling and featur e selection The combined dataset has only 486 records, which is too small to aid in de v eloping a reliable ma chine learning predicti v e model. T o increase the dataset size, a resampling technique called bootstrapping w as used to generate random samples of the dataset. One s ignicant benet of this technique is its simplicity , e xibility , rob ustness, and compatibility with e v ery data type. The technique w orks by taking a small-sized dataset and simulating samples for the dataset using the already e xisting data points. It performs simulations or formation of ne w samples by repeatedly taking random samples from the kno wn data and then replaces some samples, producing ne w instances. F or this paper , bootstrapping w as used to generate 1000 ne w datapoint samples, which were added to the already e xisting 486 samples formed from the combination of datasets. This RF algorithm pro vided by Sklearn w as used to select features that inuence the ALSFRS-R scores for patients in this paper [1]. RF w as us ed to perform feature selection due to its ability to perform intrinsic ranking of features within a da taset before making a selection [22]. The RF algorithm can also handle data with high Sensor -based pr ediction of ALS pr o gr ession: e xploring PHI and ... (Chib uzor Chukwuemeka Ok er e) Evaluation Warning : The document was created with Spire.PDF for Python.
840 ❒ ISSN: 2502-4752 dimensionality , such as in the dataset used. Finally , the dataset w as split into training sets and testing sets, for which 90% of the data w as used for training and 10% for testing. 2.6. Modeling The algorithms used in this paper include RF , Bagging Re gressor , XGBoost, MLP , FFNN, RNN and LSTM [1]. The RF , Bagging Re gressor , and XGBoost are ensemble algorithms, the M LP and FFNN are feed- forw ard algorithms, while the RNN and LSTM are RNN algorithms that w ork by propag ating backw ards to update weights of neural netw orks to reduce the error of the de v eloped model [23]. After the pre-processed steps described earlier , the split data w as used to train and test each of the predicti v e models. Ensemble algorithms were selected due to their ability to impro v e the performance of weak er algorithms, such as DT and Logistic re gression, and also, for their ability to ef ciently handle small or lar ge v olumes of dat a [16]. DNN algorithms were selected due to their ability to mimic the human brain during data p r ocessing and decision- making. 2.7. Ev aluation All de v eloped models were e v aluated using the RMSE, MAE and R-squared [7]. MAE = 1 n n X i =1 | y i − ˆ y i | (1) RMSE = v u u t 1 n n X i =1 ( y i − ˆ y i ) 2 (2) Where: y i = actual v alue y i = predicted v alue n = sample size The RMSE, MAE, and R-squared v alues for each de v eloped model were obtained and compared with other de v eloped models in the e xperiments. The model with the least error v alues and the highest R-squared v alue w as then identied as the best-performing model. 3. RESUL TS AND DISCUSSION 3.1. Description of experiments A total of 3 Experiments were used in this paper (T able 2 for description). T able 2. Summary of model de v elopment Experiment Experiment name Description Experiment 1 Features 50 T op 50 features were selected by RF and used for model de v elopment without feature engineering. Experiment 2 Features 29 Feature engineering w as done on the 50 features, and 29 features were formed. These 29 engineered features were then used for model de v elopment. Experiment 3 Features 25 Further feature engineering w as performed on Experiment 2. F our features (FVC, weight, BMI, and calories) were combined to form the PHI feature, which w as then used to impro v e the performance of the de v eloped models. 3.2. Results The e xperiments conducted in this paper compare the performance of the 3 dif ferent e xperiments with the e xclusion and inclusion of the ne wly formed PHI feature to e xamine the impact of PHI and feature engi- neering on model performance. T able 3 sho ws the performances of models de v el oped across the 3 e xperiments, and sho wing the performances of models when the PHI feature w as formed. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 835–845 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 841 T able 3. Performance comparison of dif ferent models using MAE, RMSE, and R 2 metrics Model MAE MAE MAE RMSE RMSE RMSE R 2 R 2 R 2 (Base) (Eng) (PHI) (Base ) (PHI) (Eng) (Base) (Eng) (PHI) MLP 0.258 0.174 0.262 0.546 0.442 0.552 0.633 0.757 0.651 FFNN 0.260 0.177 0.251 0.557 0.450 0.535 0.577 0.741 0.667 RNN 0.299 0.294 0.266 0.597 0.572 0.550 0.541 0.555 0.650 LSTM 0.437 0.634 0.558 0.718 0.946 0.848 0.082 -1.523 -0.195 RF 0.0666 0.0537 0.0374 0.291 0.253 0.210 0.932 0.929 0.943 Bagging 0.0889 0.0682 0.0486 0.325 0.286 0.247 0.907 0.910 0.924 XGBoost 0.0425 0.0246 0.0095 0.232 0.180 0.118 0.957 0.966 0.983 The impact of feature engineering and the creation of the PHI feature clearly impro v ed the perfor - mance of the de v eloped models. As sho wn in T able 3, all model results were lo wer . When comparing, it’ s e vident that feature engineering had a strong positi v e ef fect on model performance. F or e xample, the XGBoost model had an R-squared v alue of 95.7%, which increas ed to 96.6% after feature engineering. This sho ws that the applied feature engineering signicantly boosted model performance [1], [4], [24]. Going a step further , the introduction of the PHI feature a combination of FVC, BMI, weight, and calories led to e v en better results. T able 3 sho ws that all models performed better after this feature w as added. These ndings support the vie w of Li et al. [25], who stated that feature engineering can greatly impro v e machine learning model performance. XGBoost, which w as the best-performing model, impro v ed from MAE = 0.0178, RMSE = 0.1603, and R- squared = 0.9776 in the rst e xperiment to MSE = 0.0095, RMSE = 0.1182, and R-squared = 0.9828 in the third e xperiment. This clearly sho ws that the PHI feature contrib uted to better model results, outperforming pre vious w ork by Ok ere, Thuma and Mosweun yane [9], where the best XGBoost model had MAE = 0.4139 and RMSE = 0.7472. Additionally , DNN models, although po werful as noted by P ancotti et al. [1], did not perform better than ensemble models. This supports the ndings of Anani et al. [10] who identied ensemble models 1 as more reliable for ALS predicti on. T able 3 sho ws the R MSE trend line of the best performing mod- els from Baseline to PHI. F or e xample, XGBoost achie v ed the best performance with lo w error v alues (MAE = 0.0095, RMSE = 0.1182, R-squared = 0.9828), while the best-performing deep learning model, FFNN, only achie v ed MAE = 0.2511, RMSE = 0.5345, and R-squared = 0.6667. Figure 1 sho ws a consistent reduction in RMSE across all ensemble models as feature engineering is progressi v ely applied. Figure 1. RMSE trends of ensemble models across feature sets F or RF and Bagging Re gressor , RMSE decreases steadily from baseline to the engineered and PHI feature sets , indicating impro v ed prediction accurac y with rened features. XGBoost e xhibits the greatest impro v ement, with RMSE dropping rapidly when P HI is introduced. This trend demonstrates that PHI con- Sensor -based pr ediction of ALS pr o gr ession: e xploring PHI and ... (Chib uzor Chukwuemeka Ok er e) Evaluation Warning : The document was created with Spire.PDF for Python.
842 ❒ ISSN: 2502-4752 sistently enhances model performance across dif ferent ensemble methods, conrming that the observ ed g ains are ef fecti v e in impro ving the representational quality of the proposed feature set. In general, this s tudy sho ws that feature engineering can signicantly impro v e model performance. This aligns with the ndings of Gordon and Lerner [6] who said that feature engineering can reduce o v ertting and undertting. Senan et al. [21] also noted that combining related features into a single one, lik e PHI, can simplify comple x data and help models perform better by making the data easier to interpret. Ov erall, feature engineering reduced RMSE by up to 22%, while the introduction of PHI achie v ed an additional 34% reduction relati v e to the engineered baseline. Across all ensemble models, PHI consistently impro v ed predicti v e accurac y , with XGBoost achie ving the best performance (RMSE = 0.118, R² = 0.983). These results conrm that feature construction is critical for rob ust ALS progression modeling. 3.3. Statistical signicance analysis T o e v aluate whether the observ ed impro v ement introduced by the PHI is statistically rob ust, a paired permutation (sign-ip) test w as conducted on the cross-v alidation RMSE v alues. d i = RMSE baseline ,i − RMSE PHI ,i (3) where d i represents the paired RMSE dif ference for the i -th cross-v alidation fold, and RMSE baseline ,i and RMSE PHI ,i denote the errors obtained from the baseline and PHI-enhanced models, respecti v ely . Under the null h ypothesis H 0 : E [ d i ] = 0 , the signs of d i are e xchangeable. The proposed m odel incorporating PHI features achie v ed a substantial reduction in RMSE from 0.232 to 0.120, corresponding to an approximately 49% decrease in prediction err o r . T o assess the statistical rob ust- ness of this impro v ement, a paired permutation (sign-ip) test with 10,000 resamples w as performed on the cross-v alidation RMSE dif ferences. This non-parametric test e v aluates the null h ypothesis that the mean per - formance dif ference between the base line and PHI-enhanced models is zero, without relying on distrib utional assumptions. The analysis yielded a p-v alue of 0.031, indicating that the observ ed impro v ement is statistically signicant and unlik ely to ha v e occurred by chance. A permutation-based test w as selected due to the limited number of cross-v alidation folds, ensuring rob ustness under small-sample conditions. 3.4. Inter pr etability and clinical r ele v ance of PHI The PHI w as included in our study as it inte grates k e y indicators of patients’ nutritional, respiratory , and metabolic status, which are kno wn to inuence ALS progression. BMI is in v ersely associated with ALS survi v al: meta-analytic e vidence indicates that each additional BMI unit reduces the hazard of death by approx- imately 3%, and underweight patients e xhibit nearly double the survi v al hazard compared to normal-weight indi viduals [26]. W eight loss is a common and clini cally signicant feature in ALS, with studies reporting that 43–67% of pa tients e xperience weight loss at diagnosis, and greater weight loss correlates with w orse neurological function, f aster disease progression, and reduced survi v al [27], [28]. FVC, a measure of respi- ratory function, has been sho wn to predict both survi v al and disease progression: patients with baseline FVC < 75% e xhibit f aster functional decline and s horter survi v al, highlighting its prognostic v alue [29]. Finally , basal caloric e xpenditure reects metabolic status and ener gy demands, which are often altered in ALS and can inuence the trajectory of the disease and nutritional interv entions.By inte grating these clinically v alidated indicators into a single inde x, PHI captures complementary aspects of respiratory , metabolic, and nutritional health. This reduces feature redundanc y and noise while pres erving its meaning and enabling machine learning models to learn more stable and generalizable patterns. 4. CONCLUSION This paper sho ws that sensor data, combined with feature engineering, can be a po werful tool for tracking the progression of ALS. By combining features lik e weight, BMI, FVC, and calorie data into a single feature (PHI), it’ s possible to b uild reliable models that can help detect ALS early and monitor its progress. PHI relies on routinely collected clinical and sensor data, it can be seamlessly inte grated into clinical decisionsup- port systems and remote monitoring platforms. This enables real-time ALS progression tracking with minimal additional b urden on clinicians or patients. These models can support healthcare professionals in managing ALS more ef fecti v ely , sa ving v aluable time. Ho we v er , the study has som e limitations. A major issue is the quality of the data used. The dataset had man y missing v alues, which led to a smaller dataset after prepro- cessing. T o solv e this, future researchers could combine the IDPP CLEF 2024 dataset with other well-kno wn Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 835–845 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 843 datasets lik e the PR O-A CT dataset to increase the amount of data. Another opti o n is to collect more clinical data from hospitals and clinics. It’ s also important to use higher -quality data with fe wer missing v alues to impro v e the accurac y of the models. In terms of handling missing data, other techniques could be tested and compared with autoencoders to see which w orks best. Lastly , the models de v eloped in this study were not ne-tuned, which may ha v e af fected their performance. Future studies should consider using h yperparame- ter tuning met h ods lik e GridSearchCV and RandomSearchCV to boost model performance and also e xplore multi-modal ALS progression modeling by inte grating speech si gnals, imaging data, and genetic biomark ers alongside sensor -deri v ed features. The paper highlights the need to e xplore feature combinations as a means of enhancing machine learning models’ utilization in the area of ALS progress prediction. The de v elopment of such models can help to w ards the design and creation ef fecti v e prognostic tools to assist clinicians in decision making throughout ALS progression, enabling them to recommend appropriate and personalized therapeutic choices. A CKNO WLEDGMENTS The authors gratefully ackno wledge Brainterser project for the pro vision of the dataset used in thi s study . FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Chib uzor Chukwuemeka Ok ere ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Edwin Thuma ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Gontlafetse Mosweun yane ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data that support the ndings of this study are a v ailable from Brainterser project, who are the or g anisers of IDPP CLEF 2024 competition or g anizers [11]. REFERENCES [1] C. P ancotti, G. Birolo, C. Rollo, T . Sana via, B. Di Camillo, U. Manera, A. Chi ` o, and P . F ariselli, “Deep learning methods to predict amyotrophic lateral sclerosis disease progression, ” Scientic Reports , v ol. 12, no. (1):13738, Aug. 2022, doi: 10.1038/s41598-022- 17805-9. [2] S. B. Rutk o v e, “Clinical measures of disease progression in amyotrophic lateral sclerosis, ” Neur other apeutics , v ol. 12, no. 2, pp. 384–393, Apr . 2015, doi: 10.1007/s13311-014-0331-9. [3] P . Masrori and P . V an Damme, “ Amyotrophic lateral sclerosis: a clinical re vie w , ” Eur opean J ournal of Neur olo gy , v ol. 27, no. 10, pp. 1918–1929, Oct. 2020, doi: 10.1111/ene.14393. Sensor -based pr ediction of ALS pr o gr ession: e xploring PHI and ... (Chib uzor Chukwuemeka Ok er e) Evaluation Warning : The document was created with Spire.PDF for Python.
844 ❒ ISSN: 2502-4752 [4] F . G. V ieira, S. V enugopalan, A. S. Premasiri, M. McNally , A. Jansen, K. McClosk e y , M. P . Brenner , and S. Perrin, “ A machine-learning based objecti v e measure for ALS disease se v erity , ” npj Digital M edicine , v ol. 5, no. 1, p. 45, Apr . 2022, doi: 10.1038/s41746-022-00588-8. [5] J . D. Berry et al., “Design and results of a smartphone -based digital phenotyping study to quantify ALS progression, ” Annals of Clinical and T r anslational Neur olo gy , v ol. 6, no. 5, pp. 873–881, May 2019, doi: 10.1002/acn3.770, pMID: 31139685; PMCID: PMC6529832. [6] J . Gordon and B. Lerner , “Insights into amyotrophic lateral sclerosis from a machine learning perspecti v e, ” J ournal of Clinical Medicine , v ol. 8, no. 10, p. 1578, Oct. 2019, doi: 10.3390/jcm8101578. [7] H. T urabieh, A. S. Afshar , J. Statland, and X. Song, “T o w ards a machine learning empo wered prognostic model for predicting disease progression for amyotrophic lateral sclerosis, ” in AMIA Annual Symposium Pr oceedings , Jan. 2023, pp. 718–725, pMCID: PMC10785857 PMID: 38222431. [8] F . P apaiz, M. E. T . Dourado Jr ., R. A. D. M. V alentim, R. Pinto, A. H. F . de Morais, and J. P . Arrais, “Ensemble-imbalance-based classication for amyotrophic lateral sclerosis prognostic prediction: identifying short-survi v al patients at diagnosis, ” BMC Medical Informatics and Decision Making , v ol. 24, no. 80, Mar . 2024, doi: 10.1186/s12911-024-02484-5. [9] C. C. Ok ere, E. Thuma, and G. Mosweun yane, “UBCS at IDPP: Predicting patient self-assessment score from sensor data using machine learning algorithms, ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1392–1400. [10] T . Anani, J.-F . Pradat-Pe yre, F . Delbot, and P .-F . Pradat, “Machine learning and feature selection methods to predict 1-year survi v al and disease progression for amyotrophic lateral sclerosis, ” Preprint, 2024, doi: 10.21203/rs.3.rs-3888376/v1. [Online]. [11] G. Birolo et al., “Ov ervie w of iDPP@CLEF 2024: The intelligent disease progression prediction challenge, ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1312–1331. [12] E. Marinello, A. Guazzo, E. Long ato, E. T a v azzi, I. T rescato, M. V ettoretti, and B. Di Camillo, “Using wearable and en vironmental data to impro v e the prediction of amyotrophic lateral sclerosis and multiple sclerosis progression: an e xplorati v e study , ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1353–1365. [13] A. S. Martins, D. M. Amaral, E. N. Castanho, D. F . Soares, R. Branco, S. C. Madeira, and H. Aidos, “Predicting the functional rating scale and self-assessment status of ALS patients with sensor data, ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1366–1379. [14] R. Mehta, A. Pramo v , and S. V erma, “Machine learning for alsfrs -r score prediction: Making sense of the sens or data, ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1380–1391. [15] J. M. Silv a and J. Oli v eira, “BIT.U A at iDPP: Predicti v e analytics on als disease progression using sensor data with machine learning, ” in CEUR W orkshop Pr oceedings , v ol. 3740, 2024, pp. 1401–1411. [16] N. Zhou and P . M anser , “Does including machine learning predictions in als clinical trial analysis impro v e statistical po wer?” Annals of Clinical and T r anslational Neur olo gy , v ol. 7, no. 10, pp. 1756–1765, Oct. 2020, doi: 10.1002/acn3.51140. [17] J. E. P ark, J. H. Chung, K. H. Lee, and K. C. Shin, “The ef fect of body composition on pulmonary function, ” T uber culosis and Respir atory Diseases (Seoul) , v ol. 72, no. 5, pp. 433–440, 2012, doi: 10.4046/trd.2012.72.5.433. [18] E. McDool, J. Roone y , M. Roone y , S. Roone y , and M. Roone y , “Measuring health-related quality of life in amyotrophic lateral sclerosis, ” Neur olo gy , v ol. 103, no. 2, Jul. 2024. [19] R. W ak, “The importance of data e xploration and analysis in data science and machine learning projects, ” https://www .link edin.com/pulse/importance-data-e xploration-analysis-science-machine-learning-w ak-rqqlf/, 2024, accessed: Apr . 10, 2024. [20] G. G. K oc, C. Dagsuyu, A. K okangul, and F . K oc, “Ev aluation of alsfrs-r scale with fuzzy method in amyotrophic lateral sclerosis, ” Nor o Psikiyatr Ar sivi , v ol. 59, no. 1, pp. 54–62, 2021, doi: 10.29399/npa.27449. [21] E. M. Senan, I. Ab unadi, M. E. Jadha v , and S. M. F ati, “Score and correlation coef cient-based feature selection for predicting heart f ailure diagnosis by using machine learning algorithms, ” Computational and Mathematical Methods in Medicine , v ol. 2021, pp. 1–16, Dec. 2021, doi: 10.1155/2021/8500314. [22] P . Murug an, “Feed forw ard and backw ard run in deep con v olutional netw ork, ” in International Confer ence on Computer V ision and Ima g e Pr ocessing , v ol. 1, no. 2017, 2017, pp. 1–20. [23] T . Chai and R. R. Draxler , “Root mean square error (RMSE) or mean absolute error (MAE)?—ar guments ag ainst a v oiding rmse in the literature, ” Geoscientic Model De velopment , v ol. 7, pp. 1247–1250, 2014, doi: 10.5194/gmdd-7-1525-2014. [24] UCSF School, “ALS and neurode generati v e disease center, ” https://www .ucsfhealt h.or g/clinics/als-center , n.d., accessed: Jul. 10, 2025. [25] B. Li, A. G. Rossi, X. S. Y an, and L. Zheng, “Machine learning from a “uni v erse” of signals: The role of feature engineering, ” J ournal of F inancial Economics , v ol. 172, p. 104138, 2025. [Online]. A v ailable: https://www .sciencedirect.com/science/article/pii/S0304405X25001461 [26] E. Dardiotis, V . Siokas, M. Sokratous, Z. Tsouris, A.-M. Aloizou, D. Florou, M. Das tamani, A.-F . A. Mentis, and A. G. Brotis, “Body mass inde x and survi v al from amyotrophic lateral sclerosis: a meta-analysis, ” Neur olo gy: Clinical Pr actice , v ol. 8, no. 5, pp. 437–444, 2018, doi: 10.1212/CPJ.0000000000000521. [27] J.-Y . Li, X.-H. Sun, Z.-Y . Cai, D.-c. Shen, X.-Z. Y ang, M.-S. Liu, and L.-Y . Cui, “Correlation of weight and body composition with disease progression rate in patients with amyot rophic lateral sclerosis, ” Scientic Reports , v ol. 12, no. 1, p. 13292, 2022, 10.1038/s41598-022-16229-9. [28] A. Ja w aid et al. , “ A decrease in body mass inde x is associated with f aster progression of motor symptoms and shorter survi v al in ALS, ” Amyotr ophic Later al Scler osis , v ol. 11, no. 6, pp. 542–548, 2010, doi: doi.or g/10.3109/17482968.2010.482592. [29] A. Czaplinski, A. A. Y en, and S. H. Appel, “F orced vital capacity (FVC) as an indicator of survi v al and disease progression in an ALS clinic population, ” J ournal of Neur olo gy , Neur osur g ery and Psyc hiatry , v ol. 77, no. 3, pp. 390–392, 2006, doi: 10.1136/jnnp.2005.072660. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 3, June 2026: 835–845 Evaluation Warning : The document was created with Spire.PDF for Python.