Business Objective

Integrate different data sources like Novartis internal data, DDD, Takumi-San etc. to get complete mapping of Haematologists to Hospitals and Hospitals to SMZs. Cluster SMZs on the basis of the Haematologists to patient ratio and other features. Develop a dashboard with a geo plot depicting the different SMZs and their segmentation for targeting.

  Context

Due to lack of dedicated HEM clinics and ageing population, the number of patients outweighs Haematologists. This results in delay of care for high risk patients. Free-access allows less severe patients to visit specialists. Current data on the hospitals, HCPs, patients etc. exists in silos and is not streamlined.

  Business Outcome

Enable all hematology disease patients to receive optimal diagnosis and treatment by supporting hematologists to manage workload with task-sharing approach.

  Data Sources

Demographic: Population, Age, Area, Population Density, Binned Age, Projected Population
Medical: Patient Count, Inpatient, Outpatient, Doctors, Doctors/capita, Nursing Facilities, No of Clinics, General Doctors, Hematologists, Specialist Count, No of Specialists in each Therapy, etc
Sales

  Approach

• Integrate different data sources like Novartis internal data, DDD, Takumi-San etc. to get complete mapping of Haematologists to Hospitals and Hospitals to SMZs.
• Segregate variables based on Supply and Demand (Available slots vs Sales). This is done to obtain ratio later.
• Analyse the characteristics of SMZs and cluster on the basis of the Haematologists to patient ratio and other important latent features.
• Calculate (feature engineer) slope of yearly variables (population2015-16-17-18) to capture increase/decrease in trends and use them as variables.
• Iterate with different methodologies of clustering including K-Means, DBScan, OPTICS, BIRCH, Spectral Clustering, Agglomerative Clustering, and Gaussian Mixture Models among others.
• Develop a python dash based dashboard with a geo plot depicting the different SMZs and color coded segments.

  Algorithms

Algorithms: K-Means, DBScan, OPTICS, BIRCH, Spectral Clustering, Agglomerative, Gaussian Mixture Models
Validation: Silhouette Score, Davies Bouldin Score, Calinski Harabasz Score

  Project Outcome

Clustering and Implementation was done efficiently.Results were specific to target 9% of SMZ to affect 42% hematologists and helping 32% patients in Japan

  Awards

BOLD4CURE Award Novartis

  Business Objective

To develop and distribute a unified web application for Global Pricing Strategy team for Novartis Oncology that helps users decide a competitive price for a new lauch drug by comparing Pricing Analogs across the Oncology market.

  Context

Pricing a drug is the key factor during a product launch. The Global Pricing Strategy team wants to have an analytical approach to setting the right price of a to-be-launching drug backed by data and analytics. The right price will be a combination of characteristics from a set of already launched drug analogs.

  Business Outcome

• How to price the newly launched product and look into their projected sales and price evolution?
• What are the best suited analogues for the brand that could help give a directional on price ?
• Which are the characteristics that drive pricing decisions in different disease areas?

  Data Sources

IQVIA, PRICENTRIC, KANTAR, IPSOS, FDA, European Medicines Agency, ClinicalTrials.gov
Therapy/Indication: Therapy Area, Anatomy, Brief & Detailed Indication, Line of Therapy, Order of Entry, Orphan Launch, Immunotherapy, Chemotherapy, Targeted therapy
Molecule: Molecule Type, Mechanism of Action, Order of Entry
Patient Characteristics: Onset Age, Prevalance, 5yr Survival, Disease Type
Efficacy & Safety: Overall Survival (OS), Progression Free Survival (PFS), Overall Response Rate (ORR), CR, Quality of Life, Clinical Trial Type, No of Patients in Trial, Adverse Events (Grade3/4), % discontinued, % Severe Adverse Events
Approval: Phase of Approval, Fast Track/Breakthrough/Conditional/Priority/Accelerated Approval
Combination Drugs: Number of Combi Molecules, Combi Drug Name, Type of Combi Drug, Combi (Targeted therapy/chemo), Combi company, Cost
Competitor: Brand, Duration of Treatment, Chemo/Biologic, Offlabel, Launch Price

  Approach

• Derive Importance for key features that drive sales (Catboost Model)
• Filter brands based on selected country and product type
• Calculate similarity between the launch brand and existing brands in the database (Gower dissimilarity metric)
• Identify top K analogs using K Nearest Neighbors and Gower dissimilarity
• Determine the price bracket and show pricing evolution for the analogs
• Show analogs in a tabular format allowing users to select particular analogs of choice
• Show Price vs Sales Evolution similarity characteristics
• Show SHAP value based contribution of each variable to arrive at the projected price

  Algorithms

Regression Random Forest, Catboost for Feature Importance and Predicted Pricing
KNN Model with Gower Dissimilarity
SHAP values for contribution plot

  Project Outcome

Web application was designed in Python Plotly Dash framework and deployed to individual system using Docker

  Awards

COIN E-Avishkaar runners up ; Global Value Access Innovation Award

  Business Objective

To evaluate digital and non-digital promotional campaigns and find out the Return on Investment

  Context

The US Brand teams run digital promotions to HCPs and want to evaluate if a particular promotion was effective in generating sales and better reach patients.

  Business Outcome

Generate awareness about a particular brand to specialist HCPs so that more patients are served. HCP List Created > Forwarded to Brand Team > List forwarded to Vendors > Target Test Group is created > Campaign Runs (3 mo) > Engagement List Forwarded back > Run Analysis > Compare Test with Control

  Data Sources

Physican: TRx, NRx, NBRx Sales, Market Behaviour, Calls, Samples
Activity: Emails / Digital Newsletters / Targeted Website Banner Ads / Professional Service Sites - Opened, Clicked Through

  Approach

Starting from collating the data to delivering the recommendations in terms of ROI, the steps below list out the process:

• Collate the data for the particular brand. Find the market universe.
• Data is separated based on T and C and interaction variables are created. The original variables are also left in case the interaction variables turn out insignificant
• Avg Post TRx is predicted based on PreTRx+PostMkt+PostCalls+PostSamp for both the Test and Control HCPs
• The Pre Period variables are excluded as the T and C are matched basis that and would turn up correlated in the model.
• Model is run and if the variables are insignificant/negative coefficients, the interaction variables are removed and the total variable is used
• One equation is generated with both T and C variables residing in the model. The T and C variables are separated and separate equations are generated.
• Equation_T (Y = β1.X1+.....) for Test HCPs is generated using the X as Test_Average for each of the Independent Variables
• Equation_C (Y = β1.X1+.....) for Control HCPs is generated using the X as Test_Average for each of the Independent Variables
• Use Equation_T to calculate the Post Period TRx for each Test HCP
• Use Equation_C to calculate the Post Period TRx for each Test HCP
• The difference between, Equation_T - Equation_C gives us the lift for each Test HCP, essentially saying that if the Test HCPs did not recieve the campaign vs if the Test HCPs recieved the campaign
• Revenue = (Lift * #HCPs * #Months * $/TRx) || ROI = Revenue/Cost

  Algorithms

Test Control Methodology (Strata/1:1 Match)
T-test for validation
ANCOVA based regression modeling

  Project Outcome

Results were provided across a wide range of Campaigns starting from Email Marketing to App based notifications. Insights gererated were aligned with full flegded Marketing Mix Models.

  Business Objective

Predicting recommendation of NICE (National Institute of Health Care and Excellence) of drug based on drug characteristics.

  Context

NICE Approval is a key factor for drugs in EU. This tool shows information on
- what is the probability of a drug being recommended by NICE?
- how are the prediction results influenced by the drug characteristics?

  Business Outcome

• Identification of the governing factors for NICE recommendation will help understand the dependencies.
• Equipping the Global Pricing teams with the front-end simulation dashboard enabling deeper and better insights.

  Data Sources

NICE, IQVIA, PRICENTRIC, KANTAR, IPSOS, FDA, European Medicines Agency, ClinicalTrials.gov
Accelerated assessment (EMA), Average Survival in disease, Competitor Median Launch Price, Duration of treatment, HR QoL data availability, Launch Price (30 day therapy), Line of Therapy, Orphan Indication at Launch, ORR results, OS results, PFS results, SAE (%), Safety data (Grade III and IV AEs), Safety: % of patients discontinued due to SAE, RCT, Comparator Choice, Patient Access Scheme, End of life criteria, RWE data from CDF

  Approach

• >50 Brands’ base Data merged with additional information
• 100 attributes cleaned and pre-processed to generate insights
• Google What-if Analysis to train and simulate situations and generate deep dives
• Catboost base algorithm to select important features and retrain the model. Catboost classification method used to predict the training dataset and export the model results
• Dash Front end dashboard to take inputs from user based on certain characteristics of the new launch/ pending approval drug and show the outputs in terms of prediction of Approval or No Approval

  Algorithms

Classification Random Forest, Catboost
Explainer Dashboard for plots

  Project Outcome

Web application was designed in Explainer Dashboard framework and deployed to individual system using Docker

  Business Objective

To develop an algorithm based Customer Segmentation profile for alternate promotional channels.

  Context

Non Personal Promotions are digitally gererated promotions for HCPs. APC is another term for NPP.

  Business Outcome

Along with IC Tiers, which enable to segment the HCP population, Alternate Promotion Channel (APC) Segmentation enabled brands to have a mix of IC Tiers and APC Segments to leverage specific sets of HCPs to write more prescriptions.

  Data Sources

Product Writing (TRx, NRx, NBRx) Potential, Market Writing Potential, Last Prescription Date, First Prescription Date, Calls, Patient Switches, Patient Claims, Product Patients, Product Claims, Market Claims, Switches per Product Patient, Unique Products written in last 3 months, Connected Prescribers Adoption Potential, Depth Potential

  Approach

Integrate different data sources like Novartis internal data, Pricentric, IQVIA, Sales, etc. to get complete mapping of already launched products. Explore the data and derive important features for Product Analogs using Machine Learning

  Algorithms

Business Rule based

  Project Outcome

APC Segmentation was utilized for Marketing Campaigns specifically for NPP (Non Personal Promotions)

  Business Objective

To know how each individual marketing channel performed in the last year.

  Context

For a particular brand, it is important to know how each of the marketing campaigns (Rep Calls, Digital Touchpoints, Emails) worked in sync to affect the writing behaviour of a HCP. A marketing mix model evaluates all the campaigns together for a fixed period and generates a Average ROI, Marginal ROI and Value Due to.

  Business Outcome

The marketing mix model simulation results in understanding the previous campaigns and helps simulate and optimize spend allocation.

  Approach

• Gather the Sales Data (outcome) and the promotional channel data.
• Generate Adstock or lag factors
• Run the marketing mix model
• Calculate the VDT (Value Due to)
• Calculate the mROI and aROI

  Project Outcome

Web application was designed in Python Plotly Dash framework and deployed to individual system using Docker