Theme 11: Improving diagnosis, prognosis and disease progression.

Methods: 50 hospitalized patients with Amyotrophic Lateral Sclerosis (ALS) were enrolled in the study. All patients ran a regular rehabilitation treatment in the morning for five days a week. 25 of them were assigned to receive, in addition, Pet Therapy protocol, consisting of dog-assisted physiotherapy or occupational therapy, according to type of motor disability, three afternoons a week, while 25 control patients performed the traditional physiotherapy/occupational therapy treatment three afternoons a week. Each participant was evaluated before (T0) and after 2 weeks of training (T1). Motor performance and psycho-emotional state were assessed. All patients underwent the following scales: HADS (hospital anxiety and depression scale), MRC (Medical Research Council score) and ALSFRS-R (Revised-ALS functional rating scale). Furthermore, patients enrolled in the occupational therapy sub-groups performed nine hole peg test (NHPT) and strength measurement assessed by dynamometer, while patients enrolled in the physiotherapy group underwent Short Physical Performance Battery (SPPB) and Six Minutes Walking test (6MWT). Two different dogs were used in the therapy protocol, both were Swiss White Shepherds, with excellent and advanced education, with peculiar skills such as the ability to relate emotionally and dynamically to the patient. Results: By comparing pre-treatment (T0) vs posttreatment (T1) tests results, both pet-treated and control group and their respective physiotherapy and occupational therapy sub-groups have a trend, although not statistically significant, to improvement at all the scales taken into account. Level of anxiety results, evaluated by HADS, were, however, significantly reduced (p50.05) in the pettreated group compared with the control group.

Methods: 50 hospitalized patients with Amyotrophic Lateral Sclerosis (ALS) were enrolled in the study. All patients ran a regular rehabilitation treatment in the morning for five days a week. 25 of them were assigned to receive, in addition, Pet Therapy protocol, consisting of dog-assisted physiotherapy or occupational therapy, according to type of motor disability, three afternoons a week, while 25 control patients performed the traditional physiotherapy/occupational therapy treatment three afternoons a week. Each participant was evaluated before (T0) and after 2 weeks of training (T1). Motor performance and psycho-emotional state were assessed. All patients underwent the following scales: HADS (hospital anxiety and depression scale), MRC (Medical Research Council score) and ALSFRS-R (Revised-ALS functional rating scale). Furthermore, patients enrolled in the occupational therapy sub-groups performed nine hole peg test (NHPT) and strength measurement assessed by dynamometer, while patients enrolled in the physiotherapy group underwent Short Physical Performance Battery (SPPB) and Six Minutes Walking test (6MWT). Two different dogs were used in the therapy protocol, both were Swiss White Shepherds, with excellent and advanced education, with peculiar skills such as the ability to relate emotionally and dynamically to the patient.
Results: By comparing pre-treatment (T0) vs posttreatment (T1) tests results, both pet-treated and control group and their respective physiotherapy and occupational therapy sub-groups have a trend, although not statistically significant, to improvement at all the scales taken into account. Level of anxiety results, evaluated by HADS, were, however, significantly reduced (p50.05) in the pettreated group compared with the control group.
Discussion and conclusion: This is a pilot study showing that trained dogs can be an additional tool for global rehabilitation of patients affected by ALS. The present study demonstrates a significant therapeutic effect on anxiety. In conclusion, pet-assisted physiotherapy and occupational therapy have the same benefit on functional parameters compared to standard therapies, but have the advantage of being very appreciated by most patients, positively affecting mood, by reducing in particular anxiety. More studies are needed to explore indications and limits of the intervention. 1 IDP-02 The twin cities ALS research consortium: a model for regional collaboration in advancing research for people living with ALS Background: There is a recognized need to increase participation in ALS clinical research among people living with ALS (PALS). One of the potential barriers to participation among PALS living in metropolitan areas is a perceived or real lack of collaboration across ALS research centers. The Minneapolis-St Paul metropolitan area has a unique opportunity to overcome this barrier, as ALS providers in four academic centers share University of Minnesota faculty appointments and participate actively in regular regional clinical and research meetings hosted by the ALS Association Minnesota/North Dakota/ South Dakota chapter. Recognizing an opportunity to combine complementary strengths across centers, the ALS providers at the ALS Association Certified Treatment Centers of Excellence at Hennepin County Medical Center and the University of Minnesota have joined forces with the HealthPartners Neuroscience Center and the Minneapolis Veterans Affairs Medical Center to form the Twin Cities ALS Research Consortium (TCALSRC). All clinical research at TCALSRC centers endeavor to include principal-and co-investigators from at least two participating centers and, to whatever extent possible using courtesy appointments, investigators are encouraged to recruit and evaluate their own patients to all TCALSRC studies at whichever institution is contracted as the study site. Monthly meetings assure that all investigators and coordinators are up-to-date on current and planned projects and have an opportunity to share ideas.
Methods and results: The TCALSRC model encourages active investigator and PALS engagement throughout the metropolitan region, regardless of the study site, thus enhancing recruitment as well as value of the site to sponsors and consortia. Contracting sites are rotated among participating centers with a view toward equal participation, but also based upon the complementary research interests of individual principal investigators and the unique resources that might be available at certain institutions.
Discussions: The TCALSRC is currently exploring regional expansion to engage other outstanding ALS clinicians in Minnesota and the Dakotas. The TCALSRC is a model for advancing ALS local and regional clinical research, with several centers committed to advancing ALS research and care. Using past information about a patient, a physician can make more reliable future decisions regarding that patient. This is usually done by incorporating old with new information, giving the latter a higher weight. In this research, we use machine learning tools that mimic the physician's line of reasoning and action.
Objectives: Anytime accurate prediction of a future ALS functional rating scale (ALSFRS) a measure that reflects the patient medical condition can help both physicians and their patients to adjust the treatment and living environment of the patients along the disease period.
Methods: Using data from several clinic visits of a patient in the PROACT database and a machinelearning algorithm that combines past information about this patient with new information, we produce an accurate prediction of the patient's future state. The relevance of past information to the prediction dynamically adjusts the algorithm during training according to the level of contribution of this information to the prediction. In contrast to other machine learning models, our algorithm allows incorporation of dynamic variables, such as laboratory test results and vital signs, into the prediction.
Results: Data of 1,195 patients were used for training and testing the algorithm. Using information of all clinic visits of a patient within a period of 90 days, we predicted the ALSFRS value of this patient in a future visit, at least 270 days after the last observed visit. By comparing ALSFRS prediction values with the real ones for an examined visit, we demonstrate that our model achieves a higher accuracy than a model learned using another state-of-the-art prediction algorithm. The difference in performance between the two is statistically significant (p-value lower than 1e-06). Even when changing the prediction range, ie. the period between the last observed visit and the predicted visit, the model maintained its superiority.
Drilling down into the results, we can determine that, due to their large number in the patient population, the prediction accuracy of patients whose disease is not very progressive is even higher than the average accuracy measured over the whole population.
Discussion and conclusion: Our suggested framework for predicting the ALS disease state is analogous to that practiced by physicians and is very accurate and, thus, may easily be adopted by the medical community. IDP-04 Displaced reality: the challenges of creating an ALS clinical study in which all data collection takes place in the patient's home Background: Most ALS therapeutic clinical trials are organized in a fairly standard fashion; patients are screened and enrolled at an ALS center and return regularly for efficacy and safety evaluations. Efficacy evaluations occur every 23 months and include measures of physical function, questionnaires and blood tests. These visits are often supplemented by phone calls or emails. The total number of visits is usually less than eight over a 1-year period.
We sought to develop a new paradigm of clinical trial organization in which data collection on outcomes occurred at home either by the patient or a dedicated caregiver on a very frequent basis using simplified versions of commonly used outcome measures. Our approach differs from other groups developing web applications in which data such as activity is monitored without specific actions required by the subject.
Objectives: To describe the challenges of setting up an athome clinical study in which all the patient components, from pre-screening to consenting to data collection, are completed remotely.

Methods:
We developed an Internet-based multi-component clinical study to assess ALS progression on a day-today basis. Recruitment is achieved mainly through Internet-based advertisements. A website (ALS-at-home. org) has been developed which contains a set of prescreening questions. Upon successful completion of the pre-screening questions, the patient is contacted by a study liaison who further describes the study and arranges for medical records to be sent. After confirmation of eligibility, patients are consented via an online webinar. Subjects are then shipped devices that interact with their smart phones to allow them to collect data on speech performance, handgrip strength, multi-muscle impedance, vital capacity, functional status and activity. Ongoing compliance is monitored daily by study personnel and an automated reminder system ensures subject engagement.
Results and discussion: A variety of challenges have been identified that required a series of innovations. These included: 1) Ensuring complete anonymity during the consenting process and webinars via the use of a global unique identification (GUID) number; 2) The development of web based consenting and training videos; 3) The implementation of iOS and Android ALS-at-home apps that synced data from device-specific apps to a research electronic data capture (REDCap) database; 4) Syncing of the REDCap database with email servers to send automatic reminders to participants if they forget to collect data; and 5) Ensuring consistent performance of web and email servers to send reminders in a timely fashion.

Conclusions:
The development of a completely remote clinical study is feasible. Our work identifies a number of potential challenges to successfully launching such an investigation and straightforward steps to solving them. IDP-05 NeuroGUIDization of PALS population as a necessary condition for patient-centric research and care Background: Medical research relies on the multitude of data accumulated from various sources such as health records, clinical trials and self-reported patient information, which needs to be aggregated and records for individual patients must be linked across data sources. However, Health Insurance Portability and Accountability Act (HIPAA) regulations present significant difficulties with regard to sharing the protected health information (PHI) preventing from using PHI or any directly derived information for patient identification.

Acknowledgments
Methods: The Neurological Global Unique Patient Identifier (NeuroGUID) technology and platform make it possible to link separate datasets into a coherent harmonized data-sharing environment, while maintaining HIPAA compliance. To achieve this goal, a domainspecific central authority for generating NeuroGUIDs suitable for inclusion in de-identified datasets is set up by the Neurological Clinical Research Institute. The advantages of using NeuroGUIDs are: (a) PHI does not leave the client computer; (b) The generated ID is a random string that is not derived from the PHI; (c) IDs are centrally generated and available for use in multiple applications and on various platforms (d) Independently produced datasets can be linked together; (e) In the case of insufficient patient information for a successful NeuroGUID generation, a pseudo-ID may be generated and later be replace with a real NeuroGUID, which is advantageous for legacy datasets integration; (f) Generated NeuroGUIDs are unique yet untraceable back to the patient; and (g) Recently introduced NeuroTOKEN generation capability allows one to use unique identifiers within datasets that may contain PHI data, such as EHR/EMR systems, mobile apps, etc., ie. in any environment where the association of PHI with NeuroGUIDs should be prevented.
Results: NeuroGUIDs are widely accepted among researchers in rare neurological diseases, especially in ALS/MND. NeuroGUIDs and their derivatives (NeuroTOKENs) are utilized to connect clinical and research data to biospecimen collections (embedded into bar-coded labels on biofluids and post-mortem tissues), images (introduced into image headers), WGS files, cell lines, EHRs and mobile apps.
Conclusions: NeuroGUID technology is uniquely suitable for use with de-identified datasets. It links biosamples and images with clinical data and electronic health records. Clinicians and researchers utilize this technology, which facilitates international scientific collaboration. Background: The Pooled Resource Open-access ALS Clinical Trials (PRO-ACT) platform and database houses the largest harmonized dataset from 23 completed clinical trials in ALS (10,724 subjects). PRO-ACTÔ approach proved its efficiency and serves as a de-facto reference knowledge base with tens of published papers, developed models and performed analyses to its credit. Still, it contains a lot of information that has not been yet utilized.
Methods: There are almost 112K records of concomitant medications (ConMeds) in PRO-ACT, which we converted to the standard WHO-supported dictionary, identified active ingredients in those medications and analyzed ALSFRS-(R) slopes as follows: 1) Converted 111,848 records per WHO-DRUG dictionary; 2) Generated list of drug name and their ingredients from WHO-DRUG; 3) Converted drug names in PRO-ACT to drug ingredients, (a) ConMed records in PRO-ACT: 111,848, (b) Unique ConMeds in PRO-ACT: 6,569, (c) ConMed ingredients records in PRO-ACT: 377,163; 4) Created list of ingredients present in 25+ patient ConMed records, (a) Unique ingredients in PRO-ACT (25+ patient records): 964; 5) Created list of records for subjects with more than one ALSFRS/ALSFRS-R record: (a) Subjects: 6,565, (b) Records: 60,288; and 6) Generated matrix of ingredients present per subject in ConMeds records: (a) Average number of ingredients per subject: 27. Empirical Bayes estimates of ALSFRS-R slopes and their standard errors were obtained from a random-slopes model that included 3,504 participants who had at least 2 ALSFRS-R observations. Estimated participant-specific slopes, weighted by the inverse of their squared standard error, were regressed on known predictors of differential rates of progression (baseline ALSFRS-R, bulbar onset, age at symptom onset and symptom to diagnosis lag) and indicators for baseline use of 964 distinct medical ingredients. The data set was split 1:1:1 to training, validation and test sets. Least-angle regression was used to identify medical ingredients predictive of differential rates of ALS progression, based on reductions in averaged squared error of predictions in the validation data set.
Results: Results of meta-analysis for top 20 medical ingredients identified to potentially influence disease progression slopes will be presented.
Conclusions: With more clinical trials datasets becoming available, standard methods and approaches shall be developed to rapidly screen updated dataset for new discoveries. Data from clinics and natural history studies may improve results. IDP-07 The ALS stratification challenge: using big data and predictive computer models to identify clinically significant ALS patients sub-populations Background: One of the biggest challenges in ALS treatment and research is the disease's heterogeneity: ALS patients have widely different patterns of disease manifestation, rate of progression and clinical prognosis. This heterogeneity has detrimental effects on clinical trial planning and interpretation and on attempts to uncover disease biological mechanisms. Thus, stratifying ALS patients into clinically meaningful sub-groups can be of great value for advancing the development of effective treatments and achieving better care for ALS patients.
Objectives: We aimed to use the power of state-of-the-art machine learning algorithms applied to large-scale clinical databases of ALS patients to uncover and characterize homogeneous sub-populations of ALS patients with respect to two important clinical targets: ALSFRS progression and survival.
Methods: The 2015 DREAM ALS Stratification Prize4Life Challenge was a crowdsourcing initiative that invited participants to create algorithms to identify subgroups of patients and through that improve prediction of ALS disease progression and patients' survival. To achieve a diverse dataset, we used two data sources: ALS clinical trials data from the PRO-ACT database and communitybased ALS clinical data from ALS registries in Italy and Ireland. Thus, the challenge was divided into 4 subchallenges in which participants predicted either disease progression or survival while using data from either the PRO-ACT database or ALS registries. We assessed submitted algorithms' performance against 2 baseline algorithms, using three evaluation metrics: Pearson's correlation, concordance index, and root-mean-squaredeviation.
Results: The challenge ran between JuneOctober 2015, and final submissions were made by 30 teams, with a prize of $28,000 divided equally between the 4 best performing algorithms. For almost all teams, predictions were substantially better than random and top performing teams significantly outperformed the baseline algorithms. Patient clustering gave the most accurate predictions of disease progression for the registry data and performed generally better when applied to fast progressing patients. We used challenge participants' clustering results to detect patients who were consistently grouped together. This analysis identified, for each of the four sub-challenges, small sets of discrete 'consensus clusters'. Each of these novel subgroups of patients had distinct clinical and physiological characteristics and disease outcome profiles. A few of the most distinctive features separating the clusters were: age, time from disease onset, disease progression rate and survival, mobility, bulbar function, creatinine levels and the prevalence of bacterial infection and its biomarkers. Notably, we identified similar sub-groups of patients across different data-sets and predicted outcome measures.
Conclusions: These results demonstrate the value of large datasets and crowdsourcing challenges for developing a better understanding of ALS. We showed that ALS patients can be stratified into a few consistent and clinically-significant sub-groups, which can be used for improving prediction algorithms and guiding clinical care, research and drug development efforts. Background: A predominant problem with designing clinical trials is the heterogeneity of the ALS population. Issues such as disease progression rate, pattern of progression and more vary greatly among patients so that it is often extremely difficult to reach statistically sound conclusions in clinical trials and large numbers of participants are required for these.
In the past, studies have attempted to find meaningful sub-groups among the patient population. These studies have usually concentrated on specific features or rate of disease progression.
Objectives: Our goal was to suggest a new method of stratifying ALS patients. We aimed to show that it is important to inspect disease progression patterns as characteristics of the disease and that patients can be usefully stratified with respect to such patterns.

Methods:
We suggest a feature representation based on the multi-dimensional nature of the amyotrophic lateral sclerosis functional rating scale (ALSFRS), which allows us to capture these patterns. We leverage the natural correlation between the ALSFRS items to reduce the dimensionality of the space we are searching. We represent each patient as a vector of ALSFRS grouping derivatives (ie. deterioration rates) and then identify these patterns using unsupervised learning techniques. Specifically, we apply K-means clustering in a bootstrap sampled fashion to the database (n¼2,475) and evaluate the results with the Davies-Bouldin index (an evaluation metric for clustering partitions).
Results: We show that the best stratification as measured by the Davies-Bouldin index on the PROACT data yields an interesting division into four distinct groups of patients. These groups are differentiated by the rate of disease progression (ie. fast, moderate and slow progressors) and also by the pattern of progression. Among the groups, two can only be differentiated among by looking at specific patient functions (as represented by the different ALSFRS items) and divided into those with rapid bulbar deterioration and those with rapid limb functionality deterioration. Further, to test the validity of our distinctions, we visualized the stratification in a set of 2-dimensional spaces (ie. scatter plots) and compared characteristics of the groups. We then suggest three methods for predicting a patient's future progression pattern using features from early stages of the disease and evaluate these methods on patient data. Finally, we evaluate the usefulness of the clustering to improve predictive performance on a benchmark disease progression prediction task.
Discussion and conclusions: Our work suggests two important results: 1) inspecting disease progression patterns yields a novel, interesting and useful manner of stratifying patients, and 2) this stratification can then be used to improve performance of systems that predict future disease progression rate, thus further validating the usefulness of our results. Background: The rate of ALS disease progression is incredibly heterogeneous and, because of this, traditional ALS clinical trials often require large numbers of study participants and long treatment durations for an adequately powered trial. Several recent ALS clinical trials have aimed to select more homogenous sub-groups of ALS patients for trial enrollment in order to optimize the ability to detect a treatment effect. To better understand this strategy, we used a pooled ALS clinical trial database to study the measured treatment effect of riluzole, a medication that is known to have a modest but reproducible improvement in survival for ALS, in targeted subgroups of ALS patients in comparison to the observed treatment effect in a generalized ALS population.
Hypothesis: We hypothesized that cohorts with an overrepresentation of slow progressors would be less likely to demonstrate a significant treatment effect of riluzole.
Methods: We used the PRO-ACT ALS database to develop a gradient boosting model (GBM) survival model to predict survival at time points up to 15 months into the future. The GBM survival model was built using 4,545 PRO-ACT records that contained forced vital capacity and either ALSFRS or ALSFRS-R and indicated whether the patient was or was not taking riluzole. Patients were ranked by predicted log likelihood of survival and split into either two (fast and slow) or three (fast, average and slow) progressing groups. The groups were further divided by riluzole use and the observed Kaplan-Meier survival plots were drawn. Statistical comparisons and median survivals were calculated.
Discussion: Enriched clinical trial cohorts that select for homogenous sub-groups for trial inclusion can serve as a useful tool to improve study power and clinical trial efficiency and, more importantly, can ensure that the significance of a positive treatment effect is identified when it actually exists. Using predictive algorithms to identify patients with faster rates of disease progression for inclusion into clinical trials is a promising strategy to speed up the validation of effective disease modifying therapies in ALS. Background: Factors such as disease progression rate and pattern vary greatly among ALS patients, so it is difficult to reach statistically sound conclusions in clinical trials and to assess the effect of therapies on patients.
Objectives: To improve our ability to assess treatment influence, to decrease the sample size needed for clinical trials, and to shed more light on unknown mechanisms in the disease, we suggest a machine-learning algorithm for prediction of the ALS progression rate and creation of reliable clustering, ie. grouping of patients according to their deterioration patterns.
Methods: Since patients' condition and, thus, the ALSFRS value deteriorate over time, it is vital for a prediction algorithm to model ALS in a fully-temporal approach, namely using data from as many clinic visits as possible. An implementation of a fully-temporal machinelearning algorithm for learning a graphical model of disease trajectories is presented. The model predicts a future ALSFRS value using four components that capture the complexity of the disease trajectories: (A) Population component Shared by all patients and represents a general disease progression; (B) Sub-population component Added to the population component to reflect a progression rate of patients with similar characteristics; (C) Individual component Affects the prediction by incorporating past information regarding the individual we are making the prediction for; and (D) Noise component Models events that do not stem from the disease, but affect the patient's condition.
Results: Data of 3,925 patients from the PROACT database was used for training and testing the model. The main results of our experiments are: 1) Online prediction Our model improves prediction (its prediction error decreases) as more data regarding a patient is available for the model; 2) Clustering Using the error of future prediction as a measure, patient stratification into 10 groups/clusters gave the best performance. Visualizing the clusters' disease trajectories showed a significant difference between them, which suggests that the model has managed to group the patients well; and 3) Predicting by ALSFRS items or groups Three prediction set-ups were examined: (a) Prediction of the total ALSFRS; (b) Prediction of each ALSFRS item separately; and (c) Prediction of ALSFRS for five groups of items. Using setups (b) and (c), we were able to maintain the level of accuracy of set-up (a), while giving a more informative prediction.
Discussion and conclusions: To the best of our knowledge, this is the first research study in which ALS disease progression has been modeled using a fully-temporal approach. The results show that, by using the suggested model, we are able to improve our predictions and better cluster patients. Background: Last year at this meeting we proposed a stochastic diffusion model of ALS progression. Here, we extend the model to connect the molecular/cellular level model with the patient level by mapping patterns of neuronal degradation in the spinal cord with progression patterns in clinical assessment data. Our model is based on the hypotheses of amyotrophic lateral sclerosis (ALS) progression that posits a point source origin of motor neuron death with neuroanatomic propagation either contiguously to adjacent regions or along networks via axonal and synaptic connections. Although molecular mechanisms of propagation are unknown, one leading hypothesis is a 'prion-like' spread of misfolded and aggregated proteins, including SOD1 and TDP43.
Aim: To use mathematical models combined with clinical assessment data to quantify and characterize the cellular and molecular spread of ALS in the human spinal cord.
Methods: Our mathematical model utilizes the stochastic reaction-diffusion master equation approach on a discretized human spinal cord reconstructed from magnetic resonance (MR) images. To inform and evaluate our model on human data, we will use the ALSFRS-R score progressions (eg. swallowing, walking) mapped to somatotopic regions to create progressions for our simulated ALS model trajectories. Then, by comparing these simulated trajectories to patient progressions, we infer model parameters using the approximate Bayesian computation procedure.
Results: Our model recapitulates features of the spread of ALS in the human spinal cord. Specifically, a constant spread of ALS, despite an exponential increase in the amount of misfolded protein. We observe $20% neurons remain after the front of neurons affected by the disease has passed, which qualitatively corresponds to observations from post-mortem analysis. Combined with our analysis of the PRO-ACT clinical progression data to identify rates and classifications of disease progression and the most probable patterns of cellular degeneration, our model make it possible to test (in silico) hypotheses concerning proposed mechanisms of disease progression and validate them against the progression patterns derived from patient data.
Conclusion: Future studies will extend our preliminary model of ALS propagation to better explain variability of this disease and to address open questions in the field, namely: 1) the validity of the prion-like hypothesis of ALS, as it relates to predicting topographical propagation of neurodegeneration in ALS; 2) relative importance of contiguous/network spread in phenotypic and genetic variants of ALS; 3) relative importance of parameters in determining phenotypic variability; and 4) the role of selective vulnerability of motor neurons and other cell types in the spinal cord. Finally, we anticipate our modeling framework of disease progression will be useful in predicting the phenotypic effect of putative pharmacotherapies that may inhibit production and/or accelerate degradation/clearance of abnormal proteins in ALS and possibly other neurodegenerative disorders. Background: Patient disease heterogeneity is widely believed to be a confounding factor in the analysis of ALS clinical trials. In particular, deaths and, at the other extreme, slowly progressing patients may be the root causes of the observed heterogeneity. As a step towards solving this problem, we report on several ALS predictive models.
Objective: To validate the models using the BENEFIT-ALS placebo arm data set.

Methods:
We have developed baseline and run-in gradient boosting machine regression models (GBM) for the prediction of ALSFRS-R, gross motor sub-score, fine motor sub-score, bulbar sub-score, respiratory sub-score and vital capacity using the PRO-ACT ALS database and determined the performance characteristics of each model. Internal 10-fold cross validation was performed to assure the reproducibility of the model. The BENEFIT-ALS placebo data set (ClinicalTrials.gov trial # NCT01709149) was used as a contemporary external data set to validate and assess the generalizability of the models. Criteria for the validation of the models was set a priori to be that the root-mean square difference (RMSD) using the BENEFIT-ALS data set was to be no more than 5% greater than the RMSD of the internal validation and that the model shall not exhibit signs of bias as evidenced by a bootstrap mean prediction error for the BENEFIT-ALS data set that includes zero within the 95% confidence interval.
Results: Both run-in and baseline models were validated using the BENEFIT-ALS placebo arm data set.
Conclusions: A GBM model platform capable of predicting several key ALS disease progression metrics has been developed. This platform has applicability for both patient care and drug development. Background: This is the first step of a research study that aims to develop a machine-learning-based platform against which potential biomarkers of ALS disease can be tested as surrogate markers of disease progression. Recently developed baseline and longitudinal forced vital capacity (FVC) models are used as base models for the assessment of potential prognostic ALS biomarkers. These methods have been successfully applied to the evaluation of imaging markers in Parkinson's Disease.
Objective: We hypothesize that machine-learning-based FVC models can be used as tools to assess the relative importance of potential biomarkers as predictors of ALS disease progression.
Methods: We previously reported the development of gradient boosting machine learning models (1) for the predictions of FVC at future timepoints. These models are used to evaluate the importance of potential prognostic biomarkers relative to other predictors in the model. The performance characteristics of the models will be assessed prior to and following the addition of the potential biomarkers, alone and in combination. In addition, the importance of the biomarkers relative to other predictors in the model will be determined. Markers with predictive potential will ideally improve the overall performance of the models and rank highly in relative importance.
Results: We designed a predictive modeling platform for validation of potential biomarkers predictive of FVC progression in ALS patients. The baseline FVC model was used to evaluate the potential diagnostic value of biomarkers and the longitudinal FVC model was used to evaluate the potential prognostic value of biomarkers.
Conclusions: A machine-learning-based model platform for validation of potential biomarkers in ALS disease was developed. Our preliminary data support the hypothesis that these models can be used to assess the diagnostic and prognostic potential of ALS biomarkers.    (1)) is an enhanced version of the ALSFRS-Revised (ALSFRS-R) developed to better assess function in advanced ALS. The ALSFRS-EX contains 3 additional items; one item to further assess bulbar, fine motor and gross motor functions, respectively. The additional items are each scored via a 04 rating as in the ALSFRS-R, with a possible maximum score of 12. Since its introduction in 2009, little work has examined the utility of ALSFRS-EX.
Objectives: To examine the utility of ALSFRS-EX versus the ALSFRS-R in persons with advanced ALS in the VA Biorepository Brain Bank (VABBB) prospective ALS cohort study.
Methods: From July 2012 to the present, persons with ALS (PALS) or their caregivers were administered the ALSFRS-R with the three additional items from the ALSFRS-EX via semi-annual telephone calls. We examined the utility of the ALSFRS-EX to extend the floor of the ALSFRS-R in those participants who attained a score of 0 on the ALSFRS-R over a 60-month observation interval.
Results: 20 of 175 PALS attained a 0 on the ALSFRS-R over the observation interval. Sixteenof these 20 cases obtained greater than 0 scores on ALSFRS-EX items. The scores on the EX items ranged from 15, with 13 PALS maintaining a greater than 0 score between 654 months. The bulbar item from the ALSFRS-EX exhibited the best utility in extending the floor of the ALSFRS-R over time.
Discussion and conclusions: We found that the ALSFRS-EX improved the assessment of advanced ALS severity; particularly in the assessment of bulbar function. The ALSFRS-EX extended the floor of the ALSFRS-R for up to 54 months in the present study. Given this, and the relatively short time to administer the three additional items in the ALSFRS-EX, we recommend the use of the ALSFRS-EX, particularly in advanced ALS in clinical and research settings. Objectives: We aimed to use the widely available smartphone mobile technology to develop an application that collects objective, detailed and frequently sampled information about patients' clinical and functional status. The collected data can be used to better characterize disease course and to develop novel digital biomarkers of disease progression.

Methods: The ALS Analyzer mobile app was launched in
November 2015 and is available as a free mobile app for Android and iOS platforms. The app includes a selfreported digital version of the ALSFRS-R questionnaire and a series of tasks that estimate patients' functional abilities in all relevant functional domains: breathing capacity tests, speech tests to evaluate dysarthria, line tracing and finger tapping tests (fine motor skills), arm gross motor test and walking test. All relevant demographic information and self-reported ALSFRS-R scores are also collected through the app. The app's tasks can be completed by patients anytime, anywhere and require none or minimal assistance. All functional data is recorded via the phone's built-in sensors with no additional devices needed, making the app readily available for use by patients all over the world with only the click of a button. A range of task-specific performance parameters are recorded, allowing the development of in-depth informative analysis schemes.
Results: Over 200 ALS patients and 300 controls used the app since its launch, roughly a quarter of them using it repeatedly. Unique algorithms were developed to analyze performance in each functional task. The accumulated data was used to find thresholds separating patients from controls and to detect different levels of dysfunction in ALS patients. IDP-16 An exploratory study to investigate the use of biotelemetry to identify markers of disease progression in subjects with amyotrophic lateral sclerosis pilot phase Background: In addition to conventional clinical examination, biotelemetric monitoring of people with Amyotrophic Lateral Sclerosis (ALS) has the potential to provide an important source of information to assess the impact of the disease on aspects of functional capacity and activities of daily living, which may be a useful outcome measures in clinical trials.
Methods: This exploratory, non-controlled, non-drug study in ALS aims to investigate novel objective measures of physical activity, night-time rest, heart rate variability (HRV) and speech. There were two study phases: a variable length Pilot Phase (completed; reported here) and a 48 week Core Study Phase (ongoing; NCT02447952).
The objectives of the Pilot Phase were to (i) test the reliability, ease of use and acceptance of a wearable sensor capable of measuring simultaneous acceleration and interbeat interval (R-R), (ii) confirm that the wireless data transfer methodology (sensor-data hub-cloud) was working correctly and (iii) optimise algorithms to identify physical activities and measure HRV. Tolerability of the technology was assessed.
Results: 5 subjects attended at least one clinical visit to perform a set of pre-defined activity reference tasks while wearing the sensor on the sternum. Subjects also wore the sensor in their routine home-life setting for $3 days after the clinical visit (home monitoring). Clinic reference task acceleration data was available for four of the five subjects and home monitoring was successful for three out of five subjects; recording periods with at least 18 hours of data varied from 13 days. The charging routine (2 hours each day) was successfully adopted by three of the five subjects. The physical activity algorithms reliably classified 'lying' and 'stationary but not lying'. However, sensitivity for prediction of 'walking' was 93% and 'going up or down stairs' 40.5%. The acceleration data was, therefore, reexamined, combining 'walking' and 'stairs' into an 'active' category. With this new classifier, all three states 'lying', 'stationary but not lying' and 'active' achieved 100% accuracy for sensitivity, specificity and accuracy. HRV data were available for four of the five subjects. Acceptable quality HRV data permitting analysis varied from 74.8100% for root mean square of the successive differences (RMSSD) analyses and 53.2100% for the ratio between low and high frequency components (LF/ HF ratio) analyses.
One of the five subjects experienced moderate skin irritation and removed the sensor $5 hours before the end of the three day home monitoring recording period.
Discussion and conclusions: Overall, the sensor was generally well tolerated. User acceptance for the new technology, combined with the successful validation of the physical activity and HRV algorithms, enabled the study to continue to the Core Study Phase which is currently ongoing.
Background: The respiratory sub-score of the ALSFRS-R assesses respiratory function through items 10 (dyspnea), 11 (orthopnea) and use of non-invasive ventilation (NIV) or tracheostomy (item 12). In the South African context, NIV usage is largely related to availability, which is dependent on socioeconomic and/or cultural circumstances.
Aim: To investigate the presence of potential bias in the respiratory sub-score introduced by non-use of NIV related to availability, resulting in an apparent flattened decline in the score.

Methods:
We reviewed the records of 103 people with ALS (PALS) who are part of a prospective longitudinal study at Tygerberg Academic Hospital, South Africa, and analyzed a sub-group of patients that presented with a FVC 580% of predicted and/or respiratory symptoms, as assessed by items 10 and 11 of the ALSFRS-R. We then compared NIV usage (54 on item 12) to respiratory symptoms and FVC at the first visit (T0), T1 (up to 6 months) and T2 (up to 12 months). ) of the initial sample were alive, 17 of whom had both criteria to start NIV, but only 6 patients were using NIV. Eleven patients met criteria for introduction of NIV, but failed to do so.

Results
Discussion: As ALSFRS-R does not make provision for non-use of NIV in patients with respiratory impairment, item 12 may not be representative of actual respiratory involvement in situations where NIV use is limited by cultural or socioeconomics factors, such as resource-constrained environments. This has the potential to lead to flattening of the decline, as measured by the ALSFRS-R, and thereby result in inaccurate assessment of progression in therapeutic trials or cohort studies. Objectives: To develop a clinical outcome assessment similar to the ALSFRS-R that more accurately demonstrates PLS disease progression.
Methods: The ALSFRS-R was used as the foundation for the PLSFRS; we preserved the levels of function in each domain of the ALSFRS-R. Two additional options were added to each question (except orthopnea and respiratory insufficiency), accounting for less overall functional change than is accounted for in the ALSFRS-R. The question regarding gastrostomy was removed. Once an initial draft of the PLSFRS was developed, a patient focus group was held to determine whether the scale was meaningful to them in regard to functional change and elicit suggestions. After these suggestions were incorporated in a second draft, 20 patients from a previous study (PLS COSMOS) were contacted for feedback. This feedback was used for the final draft of the PLSFRS.
The questions now provide functional choices focused on early symptoms. The speech section accounts for changes not detectable to others and efforts to improve speech. The salivation section includes increased swallowing and coughing. The swallowing section includes issues related to gagging and coughing. The handwriting section includes modifications to writing style. Difficultly with certain foods and method modifications were added to cutting foods and handling utensils. Dressing and hygiene includes modifications to the frequency or completion of tasks. Turning in bed and adjusting sheets now accounts for intermittent assistance. Walking includes gait changes and intermittent use of assistive devices. Climbing stairs includes exercising caution. Dyspnea now includes more strenuous activities.
Results: The PLSFRS prototype was created for use in future clinical trials. A webinar was held to train evaluators from 21 sites to administer the scale. We have started the process of validating the scale with intra-and inter-rate reliability, internal consistency, construct validity and inperson and telephone-based testre-test reliability. MGH-and UPenn-UMNB scores were correlated with ALSFRS-R and SVC cross-sectionally using Pearson's correlation, longitudinally using linear mixed effects model and with disease duration at baseline using nonparametric Spearman's. For prediction analysis, a mixed model using baseline UMNB and all ALSFRS-R and SVC collected over a mean 11-months (n¼26 and 19 subjects with baseline MGH-and UPenn-UMNB, respectively) was performed.
Discussion: The UPenn-UMNB scores correlate with functional status in ALS, as measured by ALSFRS-R; MGH-UMNB does not. This correlation is mainly driven by hyperreflexia and not spasticity or PBA. Converting MGH-UMNB reflex items (biceps, triceps, patellar, ankle, Hoffman, Babinski and Jaw jerk) to binary scale to model UPenn-UMNB improved its construct validity and the correlation with ALSFRS-R was significant. Upper motor neuron dysfunction in ALS does not change over short durations. Further evaluation in a larger sample and longer follow-up is required to better characterize longitudinal changes and prediction potential of UMNB scales in ALS.
MITOS, stage by regions involved and functions lost, respectively.
Objectives: To propose an additional empirical staging system (Fine 'til 9 or FT9) drawing upon ALSFRS-R subscore trajectories dropping to a threshold of 9 (of normal 12). To explore the applicability of these 3 staging systems to the PRO-ACT database.
Methods: Distribution of stages by each system in the PRO-ACT database was examined at initial assessment and throughout the observed course, using published algorithms to convert ALSFRS-R responses to King's and MITOS stages. Markov multistate models were employed to model distribution of stages over time, estimate risks of transition from stage to stage and estimate survival by stage.
Conclusion: FT9, an empirical staging system drawing upon ALSFRS-R sub-scores, can partition the course of ALS similar to King's system, and may have the advantage of easy applicability to retrospective data. King's as well as FT9 are sensitive to observed progression of disease in the PRO-ACT cohort, whereas MITOS, being skewed towards more advanced disease, is not. Estimated transition intensities from stage to stage may be of value for counseling and research design. Introduction: Limited literature exists on preferencebased health utilities, required for the calculation of quality-adjusted life-years (QALYs) in motor neurone disease (MND). Trials of potential MND treatments often use the Amyotrophic Lateral Sclerosis Functional Rating Scale-revised (ALSFRS-R), which provides valuable information on disease progression, allows the sorting of patients by disease stage and is a predictor of survival (1). It does not, however, provide a preference-based health utility score, required for estimating QALYs in economic evaluations for health technology assessments. The ALS Utility Index (ALSUI) can be used to derive patient preferences from the ALSFRS-R, but has not been validated or used in MND patient populations. When preference-based data have not been collected directly, it may be possible to map estimated utilities from generic measures; hitherto the validity of such an approach has never been reported for MND data.
Objectives: We undertook mapping from ALSFRS-R and ALSUI to EuroQoL EQ-5D-5L domains and utility values. Furthermore, we performed indirect mapping to the EQ-5D-5L domains using the Neuropathic Pain Scale (NPS) and Hospital Anxiety and Depression Scale for MND (HAD-MND) (2).

Methods:
We developed direct mapping models using Ordinary Least Squares (OLS) and Tobit regression techniques to estimate EQ-5D-5L utilities (based on UK tariffs) with ALSFRS-R total, domain and item scores used as explanatory variables, along with ALSUI values, using patient-level data from the UK TONiC study. To map EQ-5D-5L domains, we also used indirect mapping models using the same variables, along with the NPS and HAD-MND using multinomial logistic regression techniques. We followed published mapping guidelines and reported goodness-of-fit along with predicted values for each mapping model.

Results
: The best performing model predicting EQ-5D-5L utilities used the ALSFRS-R items as explanatory variables in an OLS regression. The mean squared error was 0.0245 and the absolute mean error was 0.1228.
Discussion: Prediction was excellent with 78% of estimated values within 0.1 of the observed EQ-5D-5L utility value. Indirect mapping using the NPS and HADS provided less predictive power than direct mapping models.
Conclusions: This is the first study to present mapping algorithms to 'crosswalk' between ALSFRS-R to EQ-5D-5L. This analysis, based on TONiC data, demonstrates that the ALSFRS-R can be used to estimate EQ-5D-5L utilities when they have not been collected directly within a trial. Background: We previously reported that the 6MWT is a valid quantitative measure of walking capacity in ambALS who ambulate without (stage I) and with (stage II) assistive devices. The 6MWT was associated with measures of lower extremity muscle strength and function in both stages and with ALSFRS-R and FVC only in stage I, but not in stage II, indicating that the 6MWT is an independent measure of lower extremity function across both stage of ambulation. We also reported that serum creatinine (SC) level, a biomarker of muscle mass, predicted change in ambulation measured by ALSFRS-R walking item score.
Objective: To determine if SC correlate with the 6MWT as a measure of ambulation.
Methods: 6MWT, SC and maximum voluntary contraction of the lower extremities (MVIC) were obtained from 160 ambALS (106 in stage I and 54 in Stage II). Pairwise Pearson correlation coefficients between SC level to 6MWT distance and MVIC were obtained.

Discussion and conclusion:
The 6MWT is an independent measure of ambulatory function in both stages of ambulation. The 6MWT provides a quantitative, simple and inexpensive outcome measure of walking capacity for early stage clinical trials in ALS which correlate with muscle mass.

Conclusions:
The ALS Dashboard is a novel tool for analyzing accumulation of disease severity within a single patient and across different patients. ALS severity in some, but not all domains, segregates differently with a higher proportion of4stage-3 Bulbar Respiratory disease in the 20092011 and 20132016 cohort compared with previously. The ALS Dashboard provides a description of ALS milestone changes that permits more precision in comparing domains involved than rates of disability measured by ALSFRS-R.
Background: Amyotrophic lateral sclerosis (ALS) is characterized by progressive muscle weakness; however, no established markers of disease progression exist. Disability and death commonly result from declining respiratory muscle function, suggesting that respiratory measures such as vital capacity (VC) should be useful in predicting disease progression. Although forced vital capacity (FVC) is more typically used when making the decision to initiate non-invasive ventilation (NIV), slow VC (SVC) has been shown to be equivalent to FVC and is easier for ALS patients to perform.
Objective: To determine correlations between SVC and symptoms as measured on the revised ALS Functional Rating Scale (ALSFRS-R) in order to evaluate the utility of SVC in clinical decision-making.
Methods: The placebo group from the EMPOWER trial (Clinicaltrials.gov identifier, NCT01281189) was analyzed in the present study. The Pearson product moment correlation coefficient (r) was used to evaluate the strength of association between percentage predicted SVC and the scores of the individual respiratory sub-domain items of the ALSFRS-R, the respiratory sub-domain score and total ALSFRS-R score, respectively. Of the 467 patients randomized to the placebo group, 453 had at least 1 postbaseline measurement of SVC and were used in this analysis.
Discussion and conclusions: SVC is significantly correlated with individual respiratory symptoms measured by the ALSFRS-R (ie. dyspnea, orthopnea and respiratory insufficiency); however, only a small percentage of the variance of SVC was accounted for by any of these individual items or the overall respiratory sub-domain score. The ALSFRS-R is often used to follow disease progression and the results of this analysis suggest that changes in SVC are associated with a considerable amount of change in ALSFRS-R over time. However, the relatively low correlations observed also indicate that it is important to do repeated pulmonary testing and directly measure pulmonary function when making decisions such as when to initiate NIV. In addition, SVC is more strongly correlated with the total ALSFRS-R score than it is with the respiratory sub-domain score, suggesting that the respiratory sub-domain may not be the best predictor of respiratory decline.
IDP-25 Changes of supine forced vital capacity is the best respiratory predictor of disease progression in ALS patients  (1), but none of them is considered to be the most appropriate for describing the disease progression in ALS. On the other hand, changes of total ALSFRS-R score over the time (DFS) is considered a better indicator of the rate of progression than the total ALSFRS-R score alone (2).
Objectives: To evaluate the relationship among different routine respiratory measures decline and DFS and the ALSFRS-R total and sub-scores declines in an ALS population, during a 6-months follow-up period.
Methods: Lung functions were assessed at the first evaluation (T0) and after 6 months (T1), with seated and supine spirometry, measurement of Peak expiratory Cough Flow (PeCF), nocturnal pulse oximetry and Arterial Blood Gases analysis (ABG). ALSFRS-R total score and its bulbar (ALSFRS-Rb) and respiratory (RofALSFRS-R) sub-scores were analyzed. DFS was calculated as: 48 À ALSFRS-R at T1/disease duration from onset to T1 (month).
Discussion and conclusions: Among the respiratory measures included in this study the FVCsu showed the best correlation with the disease progression in ALS. FVC is widely used as an indicator of the prognosis, of the need for non-invasive ventilation and for assessment of efficacy of treatments in ALS. FVCsu improves the detection of diaphragmatic weakness and was found to be associated with reduced survival in ALS.
Our data reiterate that FVCsu is superior to FVCse and suggest that it may help the clinicians to predict disease progression. Its decline is a sensitive measure of global functional deterioration over time in ALS patients.
IDP-26 Alteration of forced vital capacity is a prognostic factor for survival of amyotrophic lateral sclerosis patients Methods: A cohort study was performed at our center between January 2015 and January 2017. FVC were measured in eligible patients at initial evaluation and 1year follow-up. Survival analysis was performed using a Cox model.

Results
: 211 patients with ALS diagnosed with El Escorial criteria were enrolled. Multivariate Cox's regression analysis revealed low baseline FVC at diagnosis was independently associated with poor survival outcome, which was more significant in the bulbar onset sub-group. During follow-up, the authors identified adjusted 28% (95% CI ¼ 1247%) increased risks of death associated with each 10% decrease in FVC (p50.001).
Conclusions: The measurement of FVC at the time of diagnosis and during follow-up may contribute to identify early-stage patients with poor outcome. FVC should be used to assist neurologists to adjust the clinical management for patients. Background: Early identification of bulbar involvement in persons with ALS is critical for improving diagnosis and prognosis; however, efficacious diagnostic markers have not yet been identified. Changes in lingual biomechanics during swallowing may represent a good diagnostic target, because recent work suggests that the tongue is putatively more affected in ALS than other oral structures.
Objectives: The purpose of this study was to determine whether biomechanical changes of the tongue and jaw during swallowing, measured using 3D electromagnetic articulography, pre-date clinically identifiable symptoms of speech and swallowing impairment in persons diagnosed with ALS.
Methods: Data were collected from 19 adults diagnosed with ALS and 22 neuro-typical controls. All participants (both ALS and controls) were tolerating an unrestricted diet (FOIS ¼ 7) and produced intelligible speech (497%). There were no significant differences in age or speaking rate between groups. Participants completed a 3 mL water swallow task, during which an electromagnetic tracking device recorded biomechanical measures of the anterior and posterior regions of tongue including lingual speed, range of motion, duration, coordination and efficiency. Jaw speed and range of motion were also recorded.
Discussion and conclusions: The current findings suggest that changes in lingual and jaw kinematics during a simple water task can be detected using electromagnetic articulography prior to the onset of swallowing impairment or decline in speech intelligibility in persons with ALS. Future work aims to explore the impact of lingual impairment on swallow physiology, safety and severity.