Search

Objective To employ Artificial Intelligence to model, predict and simulate the amyotrophic lateral sclerosis (ALS) progression over time in terms of variable interactions, functional impairments, and survival. Methods We employed demographic and clinical variables, including functional scores and the utilisation of support interventions, of 3940 ALS patients from four Italian and two Israeli registers to develop a new approach based on Dynamic Bayesian Networks (DBNs) that models the ALS evolution over time, in two distinct scenarios of variable availability. The method allows to simulate patients’ disease trajectories and predict the probability of functional impairment and survival at different time points. Results DBNs explicitly represent the relationships between the variables and the pathways along which they influence the disease progression. Several notable inter-dependencies were identified and validated by comparison with literature. Moreover, the implemented tool allows the assessment of the effect of different markers on the disease course, reproducing the probabilistically expected clinical progressions. The tool shows high concordance in terms of predicted and real prognosis, assessed as time to functional impairments and survival (integral of the AU-ROC in the first 36 months between 0.80–0.93 and 0.84–0.89 for the two scenarios, respectively). Conclusions Provided only with measurements commonly collected during the first visit, our models can predict time to the loss of independence in walking, breathing, swallowing, communicating, and survival and it can be used to generate in silico patient cohorts with specific characteristics. Our tool provides a comprehensive framework to support physicians in treatment planning and clinical decision-making.

Background Clinical registers constitute an invaluable resource in the medical data-driven decision making context. Accurate machine learning and data mining approaches on these data can lead to faster diagnosis, definition of tailored interventions, and improved outcome prediction. A typical issue when implementing such approaches is the almost unavoidable presence of missing values in the collected data. In this work, we propose an imputation algorithm based on a mutual information-weighted k-nearest neighbours approach, able to handle the simultaneous presence of missing information in different types of variables. We developed and validated the method on a clinical register, constituted by the information collected over subsequent screening visits of a cohort of patients affected by amyotrophic lateral sclerosis. Methods For each subject with missing data to be imputed, we create a feature vector constituted by the information collected over his/her first three months of visits. This vector is used as sample in a k-nearest neighbours procedure, in order to select, among the other patients, the ones with the most similar temporal evolution of the disease over time. An ad hoc similarity metric was implemented for the sample comparison, capable of handling the different nature of the data, the presence of multiple missing values and include the cross-information among features captured by the mutual information statistic. Results We validated the proposed imputation method on an independent test set, comparing its performance with those of three state-of-the-art competitors, resulting in better performance. We further assessed the validity of our algorithm by comparing the performance of a survival classifier built on the data imputed with our method versus the one built on the data imputed with the best-performing competitor. Conclusions Imputation of missing data is a crucial –and often mandatory– step when working with real-world datasets. The algorithm proposed in this work could effectively impute an amyotrophic lateral sclerosis clinical dataset, by handling the temporal and the mixed-type nature of the data and by exploiting the cross-information among features. We also showed how the imputation quality can affect a machine learning task.

The presence of missing data is a common problem that affects almost all clinical datasets. Since most available data mining and machine learning algorithms require complete datasets, accurately imputing (i.e. “filling in”) the missing data is an essential step. This paper presents a methodology for the missing data imputation of longitudinal clinical data based on the integration of linear interpolation and a weighted K-Nearest Neighbours (KNN) algorithm. The Maximal Information Coefficient (MIC) values among features are employed as weights for the distance computation in the KNN algorithm in order to integrate intra- and inter-patient information. An interpolation-based imputation approach was also employed and tested both independently and in combination with the KNN algorithm. The final imputation is carried out by applying the best performing method for each feature. The methodology was validated on a dataset of clinical laboratory test results of 13 commonly measured analytes of patients in an intensive care unit (ICU) setting. The performance results are compared with those of 3D-MICE, a state-of-the-art imputation method for cross-sectional and longitudinal patient data. This work was presented in the context of the 2019 ICHI Data Analytics Challenge on Missing data Imputation (DACMI).

During the acute phase of the COVID-19 pandemic, hospitals faced a challenge to manage patients, especially those with other comorbidities and medical needs, such as cancer patients. Here, we use Process Mining to analyze real-world therapeutic pathways in a cohort of 1182 cancer patients of the Lausanne University Hospital following COVID-19 infection. The algorithm builds trees representing sequences of coarse-grained events such as Home, Hospitalization, Intensive Care and Death. The same trees can also show probability of death or time-to-event statistics in each node. We introduce a new tool, called Differential Process Mining, which enables comparison of two patient strata in each node of the tree, in terms of hits and death rate, together with a statistical significance test. We thus compare management of COVID-19 patients with an active cancer in the first vs. second COVID-19 waves to quantify hospital adaptation to the pandemic. We also compare patients having undergone systemic therapy within 1 year to the rest of the cohort to understand the impact of an active cancer and/or its treatment on COVID-19 outcome. This study demonstrates the value of Process Mining to analyze complex event-based real-world data and generate hypotheses on hospital resource management or on clinical patient care.

Background Amyotrophic Lateral Sclerosis (ALS) is a neurodegenerative disease whose spreading and progression mechanisms are still unclear. The ability to predict ALS prognosis would improve the patients’ quality of life and support clinicians in planning treatments. In this paper, we investigate ALS evolution trajectories using Process Mining (PM) techniques enriched to both easily mine processes and automatically reveal how the pathways differentiate according to patients’ characteristics. Methods We consider data collected in two distinct data sources, namely the Pooled Resource Open-Access ALS Clinical Trials (PRO-ACT) dataset and a real-world clinical register (ALS–BS) including data of patients followed up in two tertiary clinical centers of Brescia (Italy). With a focus on the functional abilities progressively impaired as the disease progresses, we use two Process Discovery methods, namely the Directly-Follows Graph and the CareFlow Miner, to mine the population disease trajectories on the PRO-ACT dataset. We characterize the impairment trajectories in terms of patterns, timing, and probabilities, and investigate the effect of some patients’ characteristics at onset on the followed paths. Finally, we perform a comparative study of the impairment trajectories mined in PRO-ACT versus ALS–BS. Results We delineate the progression pathways on PRO-ACT, identifying the predominant disabilities at different stages of the disease: for instance, 85% of patients enter the trials without disabilities, and 48% of them experience the impairment of Walking/Self-care abilities first. We then test how a spinal onset increases the risk of experiencing the loss of Walking/Self-care ability as first impairment (52% vs. 27% of patients develop it as the first impairment in the spinal vs. the bulbar cohorts, respectively), as well as how an older age at onset corresponds to a more rapid progression to death. When compared, the PRO-ACT and the ALS–BS patient populations present some similarities in terms of natural progression of the disease, as well as some differences in terms of observed trajectories plausibly due to the trial scheduling and recruitment criteria. Conclusions We exploited PM to provide an overview of the evolution scenarios of an ALS trial population and to preliminary compare it to the progression observed in a clinical cohort. Future work will focus on further improving the understanding of the disease progression mechanisms, by including additional real-world subjects as well as by extending the set of events considered in the impairment trajectories.

BACKGROUND Immune checkpoint inhibitors (ICIs) have revolutionised cancer therapy but frequently cause immune-related adverse events (irAEs). Description of late-onset and duration of irAEs in the literature is often incomplete. METHODS To investigate reporting and incidence of late-onset and long-lasting irAEs, we reviewed all registration trials leading to ICI's approval by the US FDA and/or EMA up to December 2019. We analysed real-world data from all lung cancer (LC) and melanoma (Mel) patients treated with approved ICIs at the University Hospital of Lausanne (CHUV) from 2011 to 2019. To account for the immortal time bias, we used a time-dependent analysis to assess the potential association between irAEs and overall survival (OS). RESULTS Duration of irAEs and proportion of patients with ongoing toxicities at data cut-off were not specified in 56/62 (90%) publications of ICIs registration trials. In our real-world analysis, including 437 patients (217 LC, 220 Mel), 229 (52.4%) experienced at least one grade ���2 toxicity, for a total of 318 reported irAEs, of which 112 (35.2%) were long-lasting (���6 months) and about 40% were ongoing at a median follow-up of 369 days [194-695] or patient death. The cumulative probability of irAE onset from treatment initiation was 42.8%, 51.0% and 57.3% at 6, 12 and 24 months, respectively. The rate of ongoing toxicity from the time of first toxicity onset was 42.8%, 38.4% and 35.7% at 6, 12 and 24 months. Time-dependent analysis showed no significant association between the incidence of irAEs and OS in both cohorts (log Rank p��=��0.67 and 0.19 for LC and Mel, respectively). CONCLUSIONS Late-onset and long-lasting irAEs are underreported but common events during ICIs therapy. Time-dependent survival analysis is advocated to assess their impact on OS. Real-world evidence is warranted to fully capture and characterise late-onset and long-lasting irAEs in order to implement appropriate strategies for patient surveillance and follow-up.

This work presents a tool based on a Dynamic Bayesian Network (DBN) model to simulate the progression to type 2 diabetes (T2D) onset in the ageing population. Including longitudinally collected features characterizing different aspects of the ageing process, we dynamically model the relationships among the variables and the outcome over time, obtaining a network that shows a direct joined effect of glycated hemoglobin and body mass index (BMI) on the T2D onset. Remarkably, DBNs present a broad interpretability regardless of their complexity. We also employ the model to assess the impact of modifiable risk factors on developing the disease, showing how an increased BMI leads to an augmented T2D risk.

In the age of Evidence-Based Medicine, Clinical Guidelines (CGs) are recognized to be an indispensable tool to support physicians in their daily clinical practice. Medical Informatics is expected to play a relevant role in facilitating diffusion and adoption of CGs. However, the past pioneering approaches, often fragmented in many disciplines, did not lead to solutions that are actually exploited in hospitals. Process Mining for Healthcare (PM4HC) is an emerging discipline gaining the interest of healthcare experts, and seems able to deal with many important issues in representing CGs. In this position paper, we briefly describe the story and the state-of-the-art of CGs, and the efforts and results of the past approaches of medical informatics. Then, we describe PM4HC, and we answer questions like how can PM4HC cope with this challenge? Which role does PM4HC play and which rules should be employed for the PM4HC scientific community? This research received no external funding

e15095 Background: ICIs are effective in many tumors, but can cause severe irAEs, with potential long-term impact. Description of irAEs in literature is often insufficient, especially in terms of irAEs duration, and in terms of incidence of late-onset events. Namely, duration of irAEs and proportion of patients (pts) with toxicity still ongoing at data cut-off were not specified in 52 (89.6%) of 58 publications of trials leading to ICIs approval by US FDA and/or EMA up to August 2019. Real-world data are urgently needed to integrate this information. Methods: We retrieved the medical charts of all pts affected by lung cancer (LC) and melanoma (Mel) treated consecutively with approved ICIs at the Centre Hospitalier Universitaire Vaudoise (CHUV) from Jan 2012 to Dec 2019. We collected the following data: treatment setting, ICIs received, duration of each treatment, incidence of any irAEs of grade (G) ≥ 2 (according to CTCAE v5.0), time to onset and duration of irAEs. Pts with follow-up (FU) ≤ 28 days were excluded. Results: 345 pts were included (166 LC, 179 Mel), for a total of 455 ICIs treatment lines. 33 pts (7%, all with Mel) were treated in the adjuvant setting, while all the others were treated for advanced disease. 228 pts (50%) received anti-PD1, 43 (9%) anti-PD-L1, 56 (12%) anti-CTLA-4, 100 (22%) ICI-ICI combinations and 28 (6%) ICI–chemotherapy combinations. Median FU was 369 days (range 29-2903). Median duration of treatment was 71 days (range 1-1625). 196 pts (57%) experienced at least 1 G≥2 irAE for a total of 277 irAEs (22% skin, 18% gastro-intentinal [GI], 13% pneumonitis, 12% thyroid disorders, 11% hyphophisitis and 8% hepatitis). With 122/277 (44%) irAEs still on-going at our data cut-off, median duration of irAEs was 98 days; 104/277 irAEs (37%) lasted more than 6 month, of them 89% were considered irreversible. Out of 76 pts still on ICIs treatment at 6 months, incidence of G≥2 irAE with onset after 6 months was 51% (26% pneumonitis, 26% endocrine disorders, 15% skin disorders, 8% hepatitis and GI). Out of 32 pts still on ICIs treatment at 12 months, incidence of G≥2 irAE with onset after 12 months was 41% (38% pneumonitis, 23% endocrine and 15% skin disorders, 8% hepatitis and GI). Conclusions: Our real-world analysis shows that long-lasting and late-onset irAEs represent an important issue during ICIs treatment. A better description of toxicities in clinical trials is needed to allow a proper evaluation of treatment value and the impact on patients' quality of life.

e22040 Background: Process Mining (PM) is an emerging discipline with promising potential in oncology, which offers a time-oriented perspective on the clinical events that define the patient’s (pts) path of care, from diagnosis to follow-up. PM enables us to 1) build unsupervised models that automatically discover frequent patterns and detect unexpected behaviors, and 2) show the pts flow through predefined models, allowing quantitative assessments of conformance/divergence between reality and expected behaviors or medical guidelines. Methods: We explored the potential of PM to analyze treatments in a real-world melanoma cohort of 212 advanced melanoma pts treated with tyrosine kinase inhibitors (TKI), chemotherapy, or immunotherapy (IO) since 2010. 44% of pts are BRAF mutated, 24% NRAS mutated, and 16% WT. Using PM, we characterize the sequences of treatment according to type of drugs, order and timing of consecutive lines. We use kernel density estimations, Fisher or Wilcoxon tests to assess differences between groups. As main endpoint we assess the overall survival (OS) with respect to specific patterns of care followed by the pts using Kaplan-Meier analysis with log-rank test. Results: This approach provides important self-consistency checks for data and reveals the most frequent treatment sequences, and the associated outcome. The 324 administered treatments, characterized in terms of type and occurrence, are: TKI (23%), antiCTLA4 + antiPD1 (22%), antiPD1 (21%), and antiCTLA4 (20%). The overall 3-year OS is 56.9% (95% CI: 49.2 – 65.8). We identified 50 distinct patterns, which can be further aggregated into relevant cohorts to provide real-world answers to medical questions. Among others, we probed the influence of treatment order: IO single (any BRAF status), IO → TKI, TKI → IO (see table). Preliminary results suggest that pts treated with TKI → IO had shorter OS despite balance prognostic factors to IO → TKI group. IO → TKI provides short-term benefits with 1.5-year OS similar to IO alone, but degraded 3-year OS. Conclusions: The PM approach enables fast and efficient visual inspection of the data, emphasizing treatment sequences over time and allowing stratification at different levels. More population attributes and outcomes will be included to further characterize treatment sequences and provide real-world insights into future personalized care options. [Table: see text]

Headphone rendering of nearby virtual sound sources represents to date an open issue in 3-D audio, due to a number of technical challenges and temporal requirements involved in the measurement of individual Head-Related Transfer Functions (HRTFs). In order to tackle this problem, we propose a filter model of near-field effects based on the Distance Variation Function (Kan et al., 2009). Thanks to its simple structure and low order, the model can be applied to any far-field virtual auditory display to yield a realistic and computationally efficient near-field compensation of spectral and binaural effects. The model is subjectively evaluated in two psychophysical experiments where the relative distance of pairs of virtually rendered sound sources is judged. Results show that even though sound intensity overshadows subtler near-field effects when it is available as a cue for distance, the model is capable of offering relative distance information of near lateral virtual sources when intensity cues are removed. Furthermore, performances of the model in relative distance rendering are compared to those of alternative near-field rendering methods available in the literature.