#### 5.1. Theoretical Framework for Knowledge and Micromarketing

Knowledge is defined as the set of justified beliefs that enhance a firm’s capability to take effective action [

35]. Knowledge is divided into two major areas: tacit and explicit. Tacit knowledge refers to intuitions, insights and other hunches that are not easily verbalized or communicated. This tacit knowledge plays an important role in the decision-making process, especially in decision situations without the luxury of time, because it is the primary source of problem definition and alternatives [

36]. Explicit knowledge however, refers to that which can be formally collected and expressed as data, words and software; and as a result, it can readily reach consensus and quickly diffuse throughout an organization [

36]. Researchers have found that converting tacit knowledge into explicit knowledge and integrating the two can significantly enhance a firm’s competitive position by improving organizational capability, competence and performance [

37,

38]. Moreover, knowledge integration across different functions within a firm has several areas of added values, including demonstrated improvement in decision-making quality and organizational performance [

39,

40,

41,

42].

Various studies have proven that knowledge management can be improved when visualization is integrated into the modeling process because visualization’s ability to reach consensus when working with difficulty data and can simplify complex knowledge [

43]. Furthermore, integrating domain experts’ knowledge with secondary data to derive visually agreed-upon promotional response patterns has demonstrated to be an effective technique to discern responsive physicians from non-responsive physicians, leading to construction of more accurate response functions and subsequently, improving the quality of the physician detailing strategy [

4]. Likewise, it has been proven that the accuracy of promotional response function parameters for individual physicians can improve by calibrating the parameters to reveal physician responsiveness as defined by the experts [

13,

44].

The knowledge-based analytical model used in this study is based on the assumption that optimally involving domain experts and utilizing their knowledge is significant in enhancing the quality of detailing strategy. Another key assumption is that accurate PDE weights are those that visually reveal physicians’ responsiveness by matching its pattern to the predetermined responsive patterns developed by domain experts because one cannot reveal responsiveness from physicians who are unresponsive to detail [

13]. Moreover, since PDE weights are inputs to both SOV computation and detailing planning, improving the accuracy of the weights will enhance the quality of SOV calculation and detailing planning. These benefits are expected to result in the minimization of non-value-added costs, making the sales reps more effective and therefore increasing revenue.

Micromarketing is customizing marketing plans at the individualized consumer level to optimally accommodate the individual service and promotional needs to maximize the impact of sales force resources [

45,

46,

47]. Also, similar to traditional consumers, physicians respond better to marketing messages tailored to their individual needs [

44]. Consequently, we incorporated micromarketing as part of a knowledge-based approach to be more effective than the traditionally targeting physicians at a macro level approach.

#### 5.2. Process Flow of the Knowledge-Based Analytical Approach

The process flow of this approach used in the research is shown in

Figure 1 and has provided transparency to the process, which helped the organization to buy in to this study and improve the ability to coach the key stakeholders on innovation in order to sustain the correct application of this approach.

In Step 1, the definition of responsiveness, based on the visual relationship between PDE and prescription volume over time, is constructed by working with domain experts; these experts represent Marketing, Market Research, Sales Operations, Sales, and Information Management. Each expert is at least a manager with a minimum of three years’ experience in the domain as well as familiarity with the district selected for the research. These domain experts collectively defined responsiveness based on two sets of rules and those not meeting these rules are defaulted as nonresponsive. The two rules of responsiveness, which were first introduced by Yi et al. [

4], are as follows: (1) synchronized movement for all eight quarters; (2) allow for a single quarter deviation from the synchronized movement property.

Figure 2 shows examples of responsiveness based on these two rules.

The nonresponsiveness examples demonstrate cases where there exists no, or an insufficient visible pattern of relationship between PDE and prescription volume over time and are shown in

Figure 3. Clearly, detailing alone cannot explain these physicians’ prescribing behavior.

In Step 2, a neural network (NN) model is developed using the target physicians’ historical data to identify the responsive physicians. The reason for using an NN model in this study is that it automates manually intensive activity of classifying hundreds of physicians into two categories of responsiveness based on visual patterns between PDE and respective prescription volume for eight quarters developed in Step 1. Moreover, NN models are powerful pattern recognition tools especially strong in detecting nonlinear relationships between the inputs and the outputs [

48,

49]. Although NN models’ functionality is often perceived as black box, the sponsoring firm focused on performance of classification accuracy rather than the explainability of the approach [

50].

Replicating the work done by Yi et al. [

4], this research uses a back-propagation network with 16 input nodes: 8 quarters of PDE and 8 quarters of the respective prescription volume (TRx); 1 hidden layer containing 7 neurons; and 1 binary output node (1 for responsive and 0 for nonresponsive physicians). The model was developed with 450 training samples with a known output. The NN model outperformed the logistic regression model with a predicted accuracy advantage of 81% vs. 53%. The methodology was used to compare and validate the NN model’s predictive performance is shown in Equation (3).

where

Act_{i} is the actual responsiveness of physician

i;

Pred_{i} is the predicted responsiveness for physician

i;

Abs is the absolute value function; and

n is the sample size.

In Step 3, the eight quarters of PDE and respective TRx data for physicians are prepared for Step 4, the application of NN model. In Step 5, physicians’ responsiveness is identified, with nonresponsive physicians’ data directed to Step 6 and responsive physicians’ data to Step 8.

All nonresponsive physicians entering Step 6 for the first time go through to Step 7, where the NN model interface with a nonlinear mathematical model to search for a set of PDE weights that reveal physicians’ responsiveness to detailing efforts; the nonlinear program is shown below:

where NN(

Rx_{jk},

PDE_{jk}) are trained neural network function, returning 1 if physician

j is responsive and 0 if physician

j is nonresponsive to details, based on the relationship between Rx written and PDE over eight quarters;

PDE_{jk} is the physician detail equivalent for physician

j in quarter

k;

Rx_{jk} is the total number of prescriptions written by physician

j in quarter

k;

W_{ij} is the detailing weight for the

ith sequence for physician

j;

D_{ijk} is the total number of details made from the

ith sequence to physician

j in quarter

k.

The objective function given by (4), maximizes the number of responsive physicians in the first summation while maximizing the summation of the weights in the second summation. The first summation interfaces with the trained NN model by providing to the model, the physician-level prescription data and the PDE data for all eight quarters, given by Rx_{jk} and PDE_{jk} respectively, to determine the responsiveness of the targeted physicians.

The first constraint, given by (5), defines PDE for each physician in each quarter. The set of PDE weights, for the

ith position to physician

j,

W_{ij}, is initialized to 1, 0.6 and 0.3 for detailing positions 1, 2 and 3+, respectively. Constraint (6) sets the upper limit for the weight to be one. Constraint (7) forces the weights of the preceding detailing positions to be bigger than the ensuing ones in order to reflect the inverse relationship between the detailing time and the order in which a product is detailed, as well as to limit the searching space for the nonlinear mathematical program. The last constraint given by (8), defines the non-negativity constraint for all variables.

Figure 4 illustrates how this step works by having a physician visually fitting to the nonresponsive definition with the traditional set of PDE weights but the optimization algorithm has found a new set of PDE weights to make the physician fit the definition of responsiveness; and this new set of weights replaces the traditional weights for this physician, with the physician classified as responsive.

In Step 8, information on each physician’s responsiveness and the set of respective PDE weights for each is collected and stored. For the nonresponsive physicians, PDE is determined by taking the average PDE weights from the responsive physicians for each detailing sequence.

Physician responsiveness and PDE weights are merged with physician data and the company’s resource data to formulate a nonlinear programming model in Step 9. The objective of this model is to determine the optimal plan for detailing the firm’s target physicians to maximize quarterly profit. This formulation is shown here:

where

PRF_{i}(

x) is the promotional response function for physician or decile

i, and returns expected prescription volume for

x detail in a quarter;

PDE_{i} is the physician detail equivalent for responsive physician or decile

i; R

_{j} is the total quarterly resource for the

jth position details;

W_{ij} is the detailing weight for responsive physician

i for detailing position

j;

D_{ij} is the total details that need to be made in position

i to responsive physician

j;

TD_{id} is total details that need to be made in position

i to nonresponsive physician in decile

d; TP

_{d} is the total physicians in decile

d; Price is the price of a single prescription of the drug; Cost

_{i} is the cost to detailing physician

i; E is the efficiency factor to account for empty efforts directed to the physicians’ offices; and

n_{j} is the number of responsive physicians for detailing sequence

j.

The objective function, given by (9) maximizes the quarterly profit from the sales force efforts. The first summation in the function computes the optimal detailing plan to generate maximum profit from the physicians who are responsive to the sales force’s detailing efforts: the promotional response function of PDE_{i} details to physician i, given by PRF_{i}(PDE_{i}), produces the number of prescriptions written by physician i; this is then multiplied by the price per prescription, Price, to arrive at revenue; the cost to detailing physician i is given by Cost(PDE_{i}/E), where E, which denotes efficiency factor and is less than one and it accounts for the empty efforts made by the sales reps; taking the difference between the revenue and cost per physician i and summing up the profit for all the responsive physicians yields the total profit generated by this group.

The second summation in the function computes the profit from the nonresponsive physicians and since there is no visually discernable response pattern, the promotional response function to detailing effort PDE_{d}, given by PRF_{d}(PDE_{d}), is derived at decile level d based on average PDE weights from the responsive physicians. This function produces the average number of prescriptions written by an average physician from decile d; multiplying the number of prescriptions by price and subtracting the cost associated with the detailing effort, again including E, gives the profit per physician from decile d; and summing for all the deciles gives the total profit generated from this group.

Constraint (10) defines PDE for physician i, based on the PDE weights found specifically for responsive physician i in detailing position j derived earlier, in Step 7. The set of weights provides the visually recognizable pattern of responsiveness for physician i, which enables the program to locate the optimal set of details for that physician.

The average PDE weights for responsive physicians defined for each detailing position is in constraint (11). Constraint (12) defines PDE for the nonresponsive physicians per decile d. Constraint (13) sets the upper limit R_{j} on the total quarterly detailing resources for the drug for each sequence j, while the non-negativity condition is set by constraint (14). This approach was first implemented in 2005 and was still in use when we were conducting this research.