represents not only a substantial financial burden, but is also likely to decrease a patient’s adherence and compliance.9 The identification of patient response profiles is therefore very relevant; however, currently, physician’s therapeutic choices regarding biologics are based only on prior experience and a few random case reports described in literature.10 Both real-life studies and RCTs suggest that several patterns of response are possible.
In this study, we aimed to assess the characteristics of those patients, whose skin symptoms rapidly improved (so called fast-responders) and who achieved at least Psoriasis Area Severity Index (PASI) >75 in the first 4 weeks of treatment. The profiles so far hypothesized had been based on the hypothesis that the variability of the biological data and the treatments that influenced that outcome were related in a linear manner.11
Furthermore, psoriasis remains a systemic disease characterized by an intricate network of relationships between genetic12 and non-genetic factors13-15; the relationship is not fully described by linear statistics, the latter having the limit to verify only postulated hypotheses.
An alternative approach to the problem of profiling these patients is now offered by the use of Artificial Neural Networks (ANNs), which allow defining a set of computerized algorithms capable of recreating and mimicking the processes of analysis and learning typical of the human brain. This approach has recently been demonstrated to be more suitable to evaluate complex non-linear phenomena, such as biological systems.9 We applied this innovative technique to psoriasis patients treated with secukinumab aiming to predict the various patterns of response at 4 weeks of treatment. Our experience taught us that it is important for understanding of our innovative methodology to stress from the beginning that this approach is neither jeopardized by a small sample size nor by outliers.9 It makes it possible to generalize results yielded from a small population.
In this study, we aimed to assess the characteristics of those patients, whose skin symptoms rapidly improved (so called fast-responders) and who achieved at least Psoriasis Area Severity Index (PASI) >75 in the first 4 weeks of treatment. The profiles so far hypothesized had been based on the hypothesis that the variability of the biological data and the treatments that influenced that outcome were related in a linear manner.11
Furthermore, psoriasis remains a systemic disease characterized by an intricate network of relationships between genetic12 and non-genetic factors13-15; the relationship is not fully described by linear statistics, the latter having the limit to verify only postulated hypotheses.
An alternative approach to the problem of profiling these patients is now offered by the use of Artificial Neural Networks (ANNs), which allow defining a set of computerized algorithms capable of recreating and mimicking the processes of analysis and learning typical of the human brain. This approach has recently been demonstrated to be more suitable to evaluate complex non-linear phenomena, such as biological systems.9 We applied this innovative technique to psoriasis patients treated with secukinumab aiming to predict the various patterns of response at 4 weeks of treatment. Our experience taught us that it is important for understanding of our innovative methodology to stress from the beginning that this approach is neither jeopardized by a small sample size nor by outliers.9 It makes it possible to generalize results yielded from a small population.
MATERIALS AND METHODS
Study Design
Patients with moderate-to-severe plaque psoriasis (Psoriasis Area Severity Index (PASI) >10) treated with secukinumab 300 mg were enrolled in a multicenter 1) Galeazzi Hospital, 2) San Donato Hospital and 3) Policlinico in Milan, Italy) prospective open label pilot study between November 2016 and May 2017 (n=23). Other enrollment criteria were age 18–65 years, negative γâ˜immune-assay (Quantiferon®), and negative serology for hepatitis B (HBV), hepatitis C (HCV) and human immunodeficiency virus (HIV). Exclusion criteria were pregnancy, present or past erythroderma, concurrent autoimmune disease, immunodeficiency, or presence of psoriatic arthritis.
At baseline, patients underwent dermatologic assessment by two board-certified dermatologists who collected medical and pharmacological history, Body Surface Area (BSA) %, Investigator’s Global Assessment (IGA), PASI, Dermatological quality of life (DLQI), ClASsification criteria for Psoriatic ARthritis (CASPAR), and joints sonography to exclude psoriatic arthritis or enthesitis. Blood sample were collected, to provide complete blood count (CBC), transaminases, and detection of anti-nuclear antibodies. Specifically, CBC encompasses red blood cells (RBC), white blood cells (WBC), Neutrophils (N), Lymphocytes (L), Platelets (P), Aspartate Aminotransaminase (AST), Alanine Aminotransferase (ALT), P/L, N/L, AST/ALT, ALT/ AST. A second evaluation was performed 4 weeks later. To the purpose of this study, patients were categorized as “fast- responders†if they reached PASI 75, and “non-responders†if they failed to reach PASI 75. All side effects were recorded and provided to AIFA (Drug Italian Agency).
Data are expressed as the median [interquartile range (IQR)] or as a percentage. Descriptive linear statistics were used to describe the dataset. Artificial neural networks analysis (ANN) was then applied to the data.
Artificial Neural Networks Analysis
Data mining with auto-contractive map
An Auto Contractive Map (Auto-CM) was used to construct complex mathematical networks and determine the order of variables within the dataset. The Auto-CM is an unsupervised ANN that applies a learning algorithm that assigns similarities or “weights†among the input variables of a unique dataset, thus creating a square matrix of similarity. The Auto-CM Neural Network always starts with the same value, resulting in a perfectly reproducible graphical representation regardless of the number of iterations. The Auto-CM algorithm works by a) transferring signals from the Input and Hidden layers; b) assigning values between the Input layer and the Hidden layer; c) transferring signals into the Output layer from the Hidden layer; d) adjusting connections between the Hidden and the Output layers.
Once the weights are assigned, the Minimum Spanning Tree algorithm (MST) is utilized to graphically represent the shortest combination to connect the variables.11,16-19 Any connections which generate a cycle are removed to simplify the graphical representation, because all biological systems exist in a state of minimal energy and the graph represents only the fundamental biologic information. This model ultimately aims to reveal hidden trends and associations between variables by creating connections which preserve non-linear associations and visually represents them.
Simply put, the Auto-CM assigns spacing or ‘spatializes’ variables based on the correlation distance or ‘closeness’
Patients with moderate-to-severe plaque psoriasis (Psoriasis Area Severity Index (PASI) >10) treated with secukinumab 300 mg were enrolled in a multicenter 1) Galeazzi Hospital, 2) San Donato Hospital and 3) Policlinico in Milan, Italy) prospective open label pilot study between November 2016 and May 2017 (n=23). Other enrollment criteria were age 18–65 years, negative γâ˜immune-assay (Quantiferon®), and negative serology for hepatitis B (HBV), hepatitis C (HCV) and human immunodeficiency virus (HIV). Exclusion criteria were pregnancy, present or past erythroderma, concurrent autoimmune disease, immunodeficiency, or presence of psoriatic arthritis.
At baseline, patients underwent dermatologic assessment by two board-certified dermatologists who collected medical and pharmacological history, Body Surface Area (BSA) %, Investigator’s Global Assessment (IGA), PASI, Dermatological quality of life (DLQI), ClASsification criteria for Psoriatic ARthritis (CASPAR), and joints sonography to exclude psoriatic arthritis or enthesitis. Blood sample were collected, to provide complete blood count (CBC), transaminases, and detection of anti-nuclear antibodies. Specifically, CBC encompasses red blood cells (RBC), white blood cells (WBC), Neutrophils (N), Lymphocytes (L), Platelets (P), Aspartate Aminotransaminase (AST), Alanine Aminotransferase (ALT), P/L, N/L, AST/ALT, ALT/ AST. A second evaluation was performed 4 weeks later. To the purpose of this study, patients were categorized as “fast- responders†if they reached PASI 75, and “non-responders†if they failed to reach PASI 75. All side effects were recorded and provided to AIFA (Drug Italian Agency).
Data are expressed as the median [interquartile range (IQR)] or as a percentage. Descriptive linear statistics were used to describe the dataset. Artificial neural networks analysis (ANN) was then applied to the data.
Artificial Neural Networks Analysis
Data mining with auto-contractive map
An Auto Contractive Map (Auto-CM) was used to construct complex mathematical networks and determine the order of variables within the dataset. The Auto-CM is an unsupervised ANN that applies a learning algorithm that assigns similarities or “weights†among the input variables of a unique dataset, thus creating a square matrix of similarity. The Auto-CM Neural Network always starts with the same value, resulting in a perfectly reproducible graphical representation regardless of the number of iterations. The Auto-CM algorithm works by a) transferring signals from the Input and Hidden layers; b) assigning values between the Input layer and the Hidden layer; c) transferring signals into the Output layer from the Hidden layer; d) adjusting connections between the Hidden and the Output layers.
Once the weights are assigned, the Minimum Spanning Tree algorithm (MST) is utilized to graphically represent the shortest combination to connect the variables.11,16-19 Any connections which generate a cycle are removed to simplify the graphical representation, because all biological systems exist in a state of minimal energy and the graph represents only the fundamental biologic information. This model ultimately aims to reveal hidden trends and associations between variables by creating connections which preserve non-linear associations and visually represents them.
Simply put, the Auto-CM assigns spacing or ‘spatializes’ variables based on the correlation distance or ‘closeness’
